add way to split xlsx for storage

Signed-off-by: Laurent Paoletti <lp@providenz.fr>

CHANGELOG.md

@@ -9,12 +9,14 @@ and this project adheres to
 ## [Unreleased]
 
 ### Changed
-- 🐛(front) optimize chat
+
 - 📦️(front) update react
 
 ### Fixed
 
-- 🐛(e2e) fix test-e2e-chronium
+- 🐛(e2e) fix test-e2e-chromium
+- 🐛(back) fix system prompt compatibility with self-hosted models #200
+- ⚰️(back) remove dead code and unused files
 
 ## [0.0.10] - 2025-12-15
 
@@ -34,6 +36,7 @@ and this project adheres to
 ## [0.0.9] - 2025-11-17
 
 ### Added
+
 - ✨(front) add code copy button
 - ✨(RAG) add generic collection RAG tools #159
 
@@ -41,7 +44,6 @@ and this project adheres to
 
 - 🔊(langfuse) enable tracing with redacted content #162
 
-
 ## [0.0.8] - 2025-11-10
 
 ### Fixed
@@ -56,28 +58,24 @@ and this project adheres to
 
 - 🔥(posthog) remove posthog middleware for async mode fix #146
 
-
 ## [0.0.7] - 2025-10-28
 
 ### Fixed
 
 - 🚑️(posthog) fix the posthog middleware for async mode #133
 
-
 ## [0.0.6] - 2025-10-28
 
 ### Fixed
 
 - 🚑️(stats) fix tracking id in upload event #130
 
-
 ## [0.0.5] - 2025-10-27
 
 ### Fixed
 
 - 🚑️(drag-drop) fix the rejection display on Safari #127
 
-
 ## [0.0.4] - 2025-10-27
 
 ### Added
@@ -94,14 +92,12 @@ and this project adheres to
 - 🐛(front) fix mobile source
 - 🐛(attachments) reject the whole drag&drop if unsupported formats #123
 
-
 ## [0.0.3] - 2025-10-21
 
 ### Fixed
 
 - 🚑️(web-search) fix missing argument in RAG backend #116
 
-
 ## [0.0.2] - 2025-10-21
 
 ### Added
@@ -111,6 +107,7 @@ and this project adheres to
 - 📈(posthog) add `sub` field to tracking #95
 
 ### Changed
+
 - 🔧(front) change links feedback tchap + settings popup
 - 🐛(front) code activation fix session end #93
 - 💬(wording) error page wording #102
@@ -118,7 +115,6 @@ and this project adheres to
 - 🐛(activation-codes) create contact in brevo before add to list #108
 - ⚗️(summarization) add system prompt to handle tool #112
 
-
 ## [0.0.1] - 2025-10-19
 
 ### Changed
@@ -141,7 +137,7 @@ and this project adheres to
 - 🎨(front) change list attachment in chat
 - 🎨(front) move emplacement for attachment
 - 🎨(ui) retour ui sources files
-- ✨(ui) fix retour global ui 
+- ✨(ui) fix retour global ui
 - 🐛(fix) broken staging css
 - 🎨(alpha) adjustment for alpha version
 - ✨(ui) delete flex message

package.json

@@ -1,6 +0,0 @@
-{
-  "dependencies": {
-    "@ai-sdk/react": "^1.2.12",
-    "@ai-sdk/ui-utils": "^1.2.11"
-  }
-}

src/backend/chat/agent_rag/document_rag_backends/albert_rag_backend.py

@@ -18,6 +18,169 @@ from chat.agent_rag.document_rag_backends.base_rag_backend import BaseRagBackend
 
 logger = logging.getLogger(__name__)
 
+# Albert API token limit for document vectorization
+# We use a conservative chunk size to stay well under the limit
+ALBERT_MAX_TOKENS = 8192
+ALBERT_CHUNK_SIZE_TOKENS = 5000  # More conservative chunk size with larger safety margin
+# Approximate tokens: ~3 characters per token (more conservative estimate for Markdown/Excel)
+# Markdown and Excel content often have more tokens per character due to formatting
+ALBERT_CHUNK_SIZE_CHARS = ALBERT_CHUNK_SIZE_TOKENS * 3
+
+
+def _estimate_tokens(content: str) -> int:
+    """
+    Estimate the number of tokens in a text string.
+    
+    Uses a conservative approximation: ~3 characters per token.
+    This is more conservative than 4 chars/token to account for:
+    - Markdown formatting (headers, lists, tables)
+    - Excel content with special characters
+    - Whitespace and punctuation
+    
+    Args:
+        content (str): The text content to estimate.
+        
+    Returns:
+        int: Estimated number of tokens.
+    """
+    return len(content) // 3
+
+
+def _chunk_content(content: str, max_chars: int = ALBERT_CHUNK_SIZE_CHARS) -> List[str]:
+    """
+    Split content into chunks that fit within Albert's token limit.
+    
+    Attempts to split at paragraph boundaries (double newlines) when possible,
+    otherwise splits at line boundaries, and finally at character boundaries.
+    Validates that each chunk is under the token limit after splitting.
+    
+    Args:
+        content (str): The content to chunk.
+        max_chars (int): Maximum characters per chunk (default: ALBERT_CHUNK_SIZE_CHARS).
+        
+    Returns:
+        list[str]: List of content chunks, each under the token limit.
+    """
+    # First check if content fits in one chunk
+    estimated_tokens = _estimate_tokens(content)
+    if estimated_tokens <= ALBERT_CHUNK_SIZE_TOKENS:
+        return [content]
+    
+    chunks = []
+    remaining = content
+    
+    while len(remaining) > 0:
+        # Check if remaining content fits in one chunk
+        remaining_tokens = _estimate_tokens(remaining)
+        if remaining_tokens <= ALBERT_CHUNK_SIZE_TOKENS:
+            if remaining.strip():
+                chunks.append(remaining.strip())
+            break
+        
+        # Need to split - find the best split point
+        # Start with max_chars but may need to reduce if token estimate is too high
+        search_limit = max_chars
+        
+        # Try to find a split point that keeps us under token limit
+        # Reduce search limit if needed to ensure token limit is respected
+        while search_limit > 100:  # Minimum chunk size
+            # Try to split at paragraph boundary (double newline)
+            split_pos = remaining.rfind("\n\n", 0, search_limit)
+            if split_pos == -1:
+                # Try to split at single newline
+                split_pos = remaining.rfind("\n", 0, search_limit)
+            if split_pos == -1:
+                # Force split at character boundary
+                split_pos = search_limit
+            
+            # Validate that this chunk is under token limit
+            chunk_candidate = remaining[:split_pos].strip()
+            if chunk_candidate:
+                chunk_tokens = _estimate_tokens(chunk_candidate)
+                if chunk_tokens <= ALBERT_CHUNK_SIZE_TOKENS:
+                    chunks.append(chunk_candidate)
+                    remaining = remaining[split_pos:].lstrip()
+                    break
+            
+            # Chunk too large, reduce search limit and try again
+            search_limit = int(search_limit * 0.8)  # Reduce by 20%
+        else:
+            # Fallback: force split at a safe size
+            # This should rarely happen, but ensures we don't get stuck
+            safe_size = min(max_chars, len(remaining))
+            chunk = remaining[:safe_size].strip()
+            if chunk:
+                chunks.append(chunk)
+            remaining = remaining[safe_size:].lstrip()
+    
+    # Validate all chunks are under limit and split further if needed
+    validated_chunks = []
+    for chunk_item in chunks:
+        chunk_tokens = _estimate_tokens(chunk_item)
+        if chunk_tokens > ALBERT_MAX_TOKENS:
+            logger.warning(
+                "Chunk still exceeds token limit (%d tokens, max: %d), forcing split further",
+                chunk_tokens,
+                ALBERT_MAX_TOKENS,
+            )
+            # Force split this chunk further using a more conservative size
+            # Use a size that ensures we stay well under the token limit
+            # Target: ~5000 tokens max per chunk (conservative)
+            max_safe_chars = ALBERT_CHUNK_SIZE_TOKENS * 3  # 6000 * 3 = 18000 chars for ~5000 tokens
+            remaining_chunk = chunk_item
+            while len(remaining_chunk) > 0:
+                remaining_tokens = _estimate_tokens(remaining_chunk)
+                if remaining_tokens <= ALBERT_CHUNK_SIZE_TOKENS:
+                    if remaining_chunk.strip():
+                        validated_chunks.append(remaining_chunk.strip())
+                    break
+                
+                # Find a safe split point
+                split_pos = min(max_safe_chars, len(remaining_chunk))
+                # Try to split at a line boundary if possible
+                line_split = remaining_chunk.rfind("\n", 0, split_pos)
+                if line_split > max_safe_chars * 0.5:  # Only use if it's not too small
+                    split_pos = line_split
+                
+                sub_chunk = remaining_chunk[:split_pos].strip()
+                if sub_chunk:
+                    sub_tokens = _estimate_tokens(sub_chunk)
+                    # Double-check this sub-chunk is safe
+                    if sub_tokens > ALBERT_MAX_TOKENS:
+                        # Still too large, use even smaller size
+                        logger.warning(
+                            "Sub-chunk still too large (%d tokens), using smaller split",
+                            sub_tokens,
+                        )
+                        split_pos = ALBERT_CHUNK_SIZE_TOKENS * 2  # 12000 chars for ~3000 tokens
+                        sub_chunk = remaining_chunk[:split_pos].strip()
+                    validated_chunks.append(sub_chunk)
+                remaining_chunk = remaining_chunk[split_pos:].lstrip()
+        else:
+            validated_chunks.append(chunk_item)
+    
+    # Final validation - ensure NO chunk exceeds the limit
+    final_chunks = []
+    for chunk in validated_chunks:
+        chunk_tokens = _estimate_tokens(chunk)
+        if chunk_tokens > ALBERT_MAX_TOKENS:
+            logger.error(
+                "CRITICAL: Chunk still exceeds limit after all splitting attempts: %d tokens",
+                chunk_tokens,
+            )
+            # Emergency split: use very conservative size
+            emergency_size = ALBERT_CHUNK_SIZE_TOKENS * 2  # 12000 chars
+            remaining = chunk
+            while len(remaining) > 0:
+                emergency_chunk = remaining[:emergency_size].strip()
+                if emergency_chunk:
+                    final_chunks.append(emergency_chunk)
+                remaining = remaining[emergency_size:].lstrip()
+        else:
+            final_chunks.append(chunk)
+    
+    return final_chunks
+
 
 class AlbertRagBackend(BaseRagBackend):  # pylint: disable=too-many-instance-attributes
     """
@@ -170,7 +333,42 @@ class AlbertRagBackend(BaseRagBackend):  # pylint: disable=too-many-instance-att
         """
         Store the document content in the Albert collection.
         This method should handle the logic to send the document content to the Albert API.
+        
+        If the document is too large (exceeds Albert's token limit), it will be automatically
+        split into multiple chunks and stored as separate documents.
 
+        Args:
+            name (str): The name of the document.
+            content (str): The content of the document in Markdown format.
+        """
+        # Check if content needs to be chunked
+        estimated_tokens = _estimate_tokens(content)
+        
+        if estimated_tokens > ALBERT_MAX_TOKENS:
+            logger.info(
+                "Document '%s' is too large (%d estimated tokens, limit: %d). "
+                "Splitting into chunks.",
+                name,
+                estimated_tokens,
+                ALBERT_MAX_TOKENS,
+            )
+            chunks = _chunk_content(content)
+            logger.info("Split document '%s' into %d chunks", name, len(chunks))
+            
+            # Store each chunk as a separate document
+            for i, chunk in enumerate(chunks, start=1):
+                chunk_name = f"{name}_part_{i}" if len(chunks) > 1 else name
+                self._store_single_document(chunk_name, chunk)
+        else:
+            # Document fits within limit, store as-is
+            self._store_single_document(name, content)
+    
+    def _store_single_document(self, name: str, content: str) -> None:
+        """
+        Store a single document chunk in the Albert collection.
+        
+        Internal method that performs the actual API call to store one document.
+        
         Args:
             name (str): The name of the document.
             content (str): The content of the document in Markdown format.
@@ -185,14 +383,68 @@ class AlbertRagBackend(BaseRagBackend):  # pylint: disable=too-many-instance-att
             },
             timeout=settings.ALBERT_API_TIMEOUT,
         )
-        logger.debug(response.json())
+        logger.debug("Stored document '%s': %s", name, response.json())
         response.raise_for_status()
 
     async def astore_document(self, name: str, content: str) -> None:
         """
         Store the document content in the Albert collection.
         This method should handle the logic to send the document content to the Albert API.
+        
+        If the document is too large (exceeds Albert's token limit), it will be automatically
+        split into multiple chunks and stored as separate documents.
 
+        Args:
+            name (str): The name of the document.
+            content (str): The content of the document in Markdown format.
+        """
+        # Check if content needs to be chunked
+        estimated_tokens = _estimate_tokens(content)
+        
+        if estimated_tokens > ALBERT_MAX_TOKENS:
+            logger.info(
+                "Document '%s' is too large (%d estimated tokens, limit: %d). "
+                "Splitting into chunks.",
+                name,
+                estimated_tokens,
+                ALBERT_MAX_TOKENS,
+            )
+            chunks = _chunk_content(content)
+            logger.info("Split document '%s' into %d chunks", name, len(chunks))
+            
+            # Validate chunks before storing
+            for i, chunk in enumerate(chunks, start=1):
+                chunk_tokens = _estimate_tokens(chunk)
+                logger.debug(
+                    "Chunk %d/%d: %d chars, ~%d tokens",
+                    i,
+                    len(chunks),
+                    len(chunk),
+                    chunk_tokens,
+                )
+                if chunk_tokens > ALBERT_MAX_TOKENS:
+                    logger.error(
+                        "Chunk %d/%d still exceeds token limit: %d tokens (max: %d)",
+                        i,
+                        len(chunks),
+                        chunk_tokens,
+                        ALBERT_MAX_TOKENS,
+                    )
+            
+            # Store each chunk as a separate document
+            for i, chunk in enumerate(chunks, start=1):
+                chunk_name = f"{name}_part_{i}" if len(chunks) > 1 else name
+                await self._astore_single_document(chunk_name, chunk)
+        else:
+            # Document fits within limit, store as-is
+            await self._astore_single_document(name, content)
+    
+    async def _astore_single_document(self, name: str, content: str) -> None:
+        """
+        Store a single document chunk in the Albert collection.
+        
+        Internal method that performs the actual API call to store one document.
+        
         Args:
             name (str): The name of the document.
             content (str): The content of the document in Markdown format.
@@ -210,7 +462,7 @@ class AlbertRagBackend(BaseRagBackend):  # pylint: disable=too-many-instance-att
                 },
                 timeout=settings.ALBERT_API_TIMEOUT,
             )
-            logger.debug(response.json())
+            logger.debug("Stored document '%s': %s", name, response.json())
             response.raise_for_status()
 
     def search(self, query, results_count: int = 4) -> RAGWebResults:

src/backend/chat/agents/base.py

@@ -190,6 +190,4 @@ class BaseAgent(Agent):
 
         _tools = [get_pydantic_tools_by_name(tool_name) for tool_name in self.configuration.tools]
 
-        super().__init__(
-            model=_model_instance, system_prompt=_system_prompt, tools=_tools, **kwargs
-        )
+        super().__init__(model=_model_instance, instructions=_system_prompt, tools=_tools, **kwargs)

src/backend/chat/agents/conversation.py

@@ -16,7 +16,6 @@ from .base import BaseAgent
 
 logger = logging.getLogger(__name__)
 
-
 MOCKED_RESPONSE = """
 # **Ode to the AI Assistant** 🤖✨
 
@@ -102,10 +101,10 @@ class ConversationAgent(BaseAgent):
         if settings.WARNING_MOCK_CONVERSATION_AGENT:
             self._model = FunctionModel(stream_function=mocked_agent_model)
 
-        @self.system_prompt
+        @self.instructions
         def add_the_date() -> str:
             """
-            Dynamic system prompt function to add the current date.
+            Dynamic instruction function to add the current date.
 
             Warning: this will always use the date in the server timezone,
             not the user's timezone...
@@ -113,9 +112,9 @@ class ConversationAgent(BaseAgent):
             _formatted_date = formats.date_format(timezone.now(), "l d/m/Y", use_l10n=False)
             return f"Today is {_formatted_date}."
 
-        @self.system_prompt
+        @self.instructions
         def enforce_response_language() -> str:
-            """Dynamic system prompt function to set the expected language to use."""
+            """Dynamic instruction function to set the expected language to use."""
             return f"Answer in {get_language_name(language).lower()}." if language else ""
 
     def get_web_search_tool_name(self) -> str | None:

src/backend/chat/clients/pydantic_ai.py

@@ -72,12 +72,16 @@ from chat.clients.pydantic_ui_message_converter import (
     ui_message_to_user_content,
 )
 from chat.mcp_servers import get_mcp_servers
+from chat.tools.data_analysis import add_data_analysis_tool
 from chat.tools.document_generic_search_rag import add_document_rag_search_tool_from_setting
 from chat.tools.document_search_rag import add_document_rag_search_tool
 from chat.tools.document_summarize import document_summarize
 from chat.vercel_ai_sdk.core import events_v4, events_v5
 from chat.vercel_ai_sdk.encoder import EventEncoder
 
+# Keep at the top of the file to avoid mocking issues
+document_store_backend = import_string(settings.RAG_DOCUMENT_SEARCH_BACKEND)
+
 logger = logging.getLogger(__name__)
 
 User = get_user_model()
@@ -148,6 +152,7 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
             deps_type=ContextDeps,
         )
         add_document_rag_search_tool_from_setting(self.conversation_agent, self.user)
+        add_data_analysis_tool(self.conversation_agent)
 
     @property
     def _stop_cache_key(self):
@@ -236,6 +241,7 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
         Parse and store input documents in the conversation's document store.
         """
         # Early external document URL rejection
+
         if any(
             not document.url.startswith("/media-key/")
             for document in documents
@@ -249,8 +255,6 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
         ):
             raise ValueError("Document URL does not belong to the conversation.")
 
-        document_store_backend = import_string(settings.RAG_DOCUMENT_SEARCH_BACKEND)
-
         document_store = document_store_backend(self.conversation.collection_id)
         if not document_store.collection_id:
             # Create a new collection for the conversation
@@ -287,7 +291,24 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
                     content=document.data,
                 )
 
-            if not document.media_type.startswith("text/"):
+            # Don't convert tabular files (CSV, Excel) to Markdown - keep originals for data_analysis tool
+            # Tabular files are already text-based or can be used directly
+            is_tabular_file = (
+                document.media_type in [
+                    "text/csv",
+                    "application/csv",
+                    "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
+                    "application/vnd.ms-excel",
+                    "application/excel",
+                ]
+                or any(
+                    document.identifier.lower().endswith(ext)
+                    for ext in [".csv", ".xlsx", ".xls", ".xlsm", ".xlsb"]
+                )
+            )
+            
+            # Only convert non-text files that are not tabular files
+            if not document.media_type.startswith("text/") and not is_tabular_file:
                 md_attachment = await models.ChatConversationAttachment.objects.acreate(
                     conversation=self.conversation,
                     uploaded_by=self.user,
@@ -420,6 +441,7 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
                     ],
                 },
             )
+
             try:
                 await self.parse_input_documents(input_documents)
             except Exception as exc:  # pylint: disable=broad-except
@@ -457,7 +479,7 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
 
         if force_web_search:
 
-            @self.conversation_agent.system_prompt
+            @self.conversation_agent.instructions
             def force_web_search_prompt() -> str:
                 """Dynamic system prompt function to force web search."""
                 return (
@@ -484,6 +506,7 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
             .aexists()
         )
 
+
         document_urls = []
         if not conversation_has_documents and not has_not_pdf_docs:
             # No documents to process
@@ -505,7 +528,7 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
                 )
 
             # Inform the model (system-level) that documents are attached and available
-            @self.conversation_agent.system_prompt
+            @self.conversation_agent.instructions
             def attached_documents_note() -> str:
                 return (
                     "[Internal context] User documents are attached to this conversation. "
@@ -518,6 +541,13 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
             async def summarize(ctx: RunContext, *args, **kwargs) -> ToolReturn:
                 """Wrap the document_summarize tool to provide context and add the tool."""
                 return await document_summarize(ctx, *args, **kwargs)
+
+        if not conversation_has_documents and not has_not_pdf_docs:
+            # No documents to process
+            pass
+        elif has_not_pdf_docs:
+            # Already handled above with RAG tool
+            pass
         else:
             conversation_documents = [
                 cd
@@ -731,7 +761,6 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
             langfuse.update_current_trace(
                 output=run.result.output if self._store_analytics else "REDACTED"
             )
-
         # Vercel finish message
         yield events_v4.FinishMessagePart(
             finish_reason=events_v4.FinishReason.STOP,

src/backend/chat/tests/agents/test_build_conversation_agent.py

@@ -27,9 +27,14 @@ def test_build_pydantic_agent_success_no_tools():
     """Test successful agent creation without tools."""
     agent = ConversationAgent(model_hrid="default-model")
     assert isinstance(agent, Agent)
+    assert agent._system_prompts == ()
+
+    instructions = agent._instructions
+    assert len(instructions) == 3
+    assert instructions[0] == "You are a helpful assistant"
+    assert instructions[1].__name__ == "add_the_date"
+    assert instructions[2].__name__ == "enforce_response_language"
 
-    assert agent._system_prompts == ("You are a helpful assistant",)
-    assert agent._instructions == []
     assert isinstance(agent.model, OpenAIChatModel)
     assert agent.model.model_name == "model-123"
     assert str(agent.model.client.base_url) == "https://api.llm.com/v1/"
@@ -37,6 +42,7 @@ def test_build_pydantic_agent_success_no_tools():
     assert agent._function_toolset.tools == {}
 
 
+@freeze_time("2025-07-25T10:36:35.297675Z")
 def test_build_pydantic_agent_with_tools(settings):
     """Test successful agent creation with tools."""
     settings.AI_AGENT_TOOLS = ["get_current_weather"]
@@ -44,8 +50,14 @@ def test_build_pydantic_agent_with_tools(settings):
     agent = ConversationAgent(model_hrid="default-model")
     assert isinstance(agent, Agent)
 
-    assert agent._system_prompts == ("You are a helpful assistant",)
-    assert agent._instructions == []
+    instructions = agent._instructions
+    assert len(instructions) == 3
+    assert instructions[0] == "You are a helpful assistant"
+    assert instructions[1].__name__ == "add_the_date"
+    assert instructions[1]() == "Today is Friday 25/07/2025."
+    assert instructions[2].__name__ == "enforce_response_language"
+    assert instructions[2]() == ""
+
     assert isinstance(agent.model, OpenAIChatModel)
     assert agent.model.model_name == "model-123"
     assert str(agent.model.client.base_url) == "https://api.llm.com/v1/"
@@ -56,21 +68,23 @@ def test_build_pydantic_agent_with_tools(settings):
 @freeze_time("2025-07-25T10:36:35.297675Z")
 def test_add_dynamic_system_prompt():
     """
-    Ensure add_the_date and enforce_response_language system prompt are registered
+    Ensure add_the_date and enforce_response_language instructions are registered
     and returns proper values.
     """
     agent = ConversationAgent(model_hrid="default-model")
 
-    assert len(agent._system_prompt_functions) == 2
+    assert len(agent._system_prompt_functions) == 0
 
-    assert agent._system_prompt_functions[0].function.__name__ == "add_the_date"
-    assert agent._system_prompt_functions[0].function() == "Today is Friday 25/07/2025."
-
-    assert agent._system_prompt_functions[1].function.__name__ == "enforce_response_language"
-    assert agent._system_prompt_functions[1].function() == ""
+    instructions = agent._instructions
+    assert len(instructions) == 3
+    assert instructions[0] == "You are a helpful assistant"
+    assert instructions[1].__name__ == "add_the_date"
+    assert instructions[1]() == "Today is Friday 25/07/2025."
+    assert instructions[2].__name__ == "enforce_response_language"
+    assert instructions[2]() == ""
 
     agent = ConversationAgent(model_hrid="default-model", language="fr-fr")
-    assert agent._system_prompt_functions[1].function() == "Answer in french."
+    assert agent._instructions[2]() == "Answer in french."
 
 
 def test_agent_get_web_search_tool_name(settings):

src/backend/chat/tests/views/chat/conversations/test_conversation.py

@@ -130,6 +130,16 @@ def test_post_conversation_data_protocol(api_client, mock_openai_stream):
 
     assert mock_openai_stream.called
 
+    # ensure instructions are merged as a system prompt
+    last_request_payload = json.loads(respx.calls.last.request.content)
+    assert last_request_payload["messages"][0] == {
+        "content": (
+            "You are a helpful test assistant :)\n\nToday is Friday 25/07/2025.\n\n"
+            "Answer in english."
+        ),
+        "role": "system",
+    }
+
     chat_conversation.refresh_from_db()
     assert chat_conversation.ui_messages == [
         {
@@ -170,29 +180,15 @@ def test_post_conversation_data_protocol(api_client, mock_openai_stream):
     )
 
     _run_id = chat_conversation.pydantic_messages[0]["run_id"]
+
     assert chat_conversation.pydantic_messages == [
         {
-            "instructions": None,
+            "instructions": (
+                "You are a helpful test assistant :)\n\n"
+                "Today is Friday 25/07/2025.\n\nAnswer in english."
+            ),
             "kind": "request",
             "parts": [
-                {
-                    "content": "You are a helpful test assistant :)",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Today is Friday 25/07/2025.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Answer in english.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
                 {
                     "content": ["Hello"],
                     "part_kind": "user-prompt",
@@ -255,6 +251,15 @@ def test_post_conversation_text_protocol(api_client, mock_openai_stream):
     assert response_content == "Hello there"
 
     assert mock_openai_stream.called
+    # ensure instructions are merged as a system prompt
+    last_request_payload = json.loads(respx.calls.last.request.content)
+    assert last_request_payload["messages"][0] == {
+        "content": (
+            "You are a helpful test assistant :)\n\nToday is Friday 25/07/2025.\n\n"
+            "Answer in english."
+        ),
+        "role": "system",
+    }
 
     chat_conversation.refresh_from_db()
     assert chat_conversation.ui_messages == [
@@ -296,29 +301,15 @@ def test_post_conversation_text_protocol(api_client, mock_openai_stream):
     )
 
     _run_id = chat_conversation.pydantic_messages[0]["run_id"]
+
     assert chat_conversation.pydantic_messages == [
         {
-            "instructions": None,
+            "instructions": (
+                "You are a helpful test assistant :)\n\n"
+                "Today is Friday 25/07/2025.\n\nAnswer in english."
+            ),
             "kind": "request",
             "parts": [
-                {
-                    "content": "You are a helpful test assistant :)",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Today is Friday 25/07/2025.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Answer in english.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
                 {
                     "content": ["Hello"],
                     "part_kind": "user-prompt",
@@ -409,11 +400,12 @@ def test_post_conversation_with_image(api_client, mock_openai_stream_image):
     # Check the exact structure expected by the AI service
     assert body["messages"] == [
         {
-            "content": "You are a helpful test assistant :)",
+            "content": (
+                "You are a helpful test assistant :)\n\nToday is Friday 25/07/2025."
+                "\n\nAnswer in english."
+            ),
             "role": "system",
         },
-        {"content": "Today is Friday 25/07/2025.", "role": "system"},
-        {"content": "Answer in english.", "role": "system"},
         {
             "content": [
                 {"text": "Hello, what do you see on this picture?", "type": "text"},
@@ -498,27 +490,12 @@ def test_post_conversation_with_image(api_client, mock_openai_stream_image):
     _run_id = chat_conversation.pydantic_messages[0]["run_id"]
     assert chat_conversation.pydantic_messages == [
         {
-            "instructions": None,
+            "instructions": (
+                "You are a helpful test assistant :)\n\n"
+                "Today is Friday 25/07/2025.\n\nAnswer in english."
+            ),
             "kind": "request",
             "parts": [
-                {
-                    "content": "You are a helpful test assistant :)",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Today is Friday 25/07/2025.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Answer in english.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
                 {
                     "content": [
                         "Hello, what do you see on this picture?",
@@ -616,11 +593,12 @@ def test_post_conversation_tool_call(api_client, mock_openai_stream_tool, settin
 
     assert body["messages"] == [
         {
-            "content": "You are a helpful test assistant :)",
+            "content": (
+                "You are a helpful test assistant :)\n\n"
+                "Today is Friday 25/07/2025.\n\nAnswer in english."
+            ),
             "role": "system",
         },
-        {"content": "Today is Friday 25/07/2025.", "role": "system"},
-        {"content": "Answer in english.", "role": "system"},
         {"content": [{"text": "Weather in Paris?", "type": "text"}], "role": "user"},
     ]
 
@@ -678,27 +656,12 @@ def test_post_conversation_tool_call(api_client, mock_openai_stream_tool, settin
     _run_id = chat_conversation.pydantic_messages[0]["run_id"]
     assert chat_conversation.pydantic_messages == [
         {
-            "instructions": None,
+            "instructions": (
+                "You are a helpful test assistant :)\n\n"
+                "Today is Friday 25/07/2025.\n\nAnswer in english."
+            ),
             "kind": "request",
             "parts": [
-                {
-                    "content": "You are a helpful test assistant :)",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Today is Friday 25/07/2025.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Answer in english.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
                 {
                     "content": ["Weather in Paris?"],
                     "part_kind": "user-prompt",
@@ -737,7 +700,10 @@ def test_post_conversation_tool_call(api_client, mock_openai_stream_tool, settin
             "run_id": _run_id,
         },
         {
-            "instructions": None,
+            "instructions": (
+                "You are a helpful test assistant :)\n\n"
+                "Today is Friday 25/07/2025.\n\nAnswer in english."
+            ),
             "kind": "request",
             "parts": [
                 {
@@ -829,11 +795,12 @@ def test_post_conversation_tool_call_fails(api_client, mock_openai_stream_tool,
 
     assert body["messages"] == [
         {
-            "content": "You are a helpful test assistant :)",
+            "content": (
+                "You are a helpful test assistant :)\n\n"
+                "Today is Friday 25/07/2025.\n\nAnswer in french."
+            ),
             "role": "system",
         },
-        {"content": "Today is Friday 25/07/2025.", "role": "system"},
-        {"content": "Answer in french.", "role": "system"},
         {"content": [{"text": "Weather in Paris?", "type": "text"}], "role": "user"},
     ]
 
@@ -891,27 +858,12 @@ def test_post_conversation_tool_call_fails(api_client, mock_openai_stream_tool,
     _run_id = chat_conversation.pydantic_messages[0]["run_id"]
     assert chat_conversation.pydantic_messages == [
         {
-            "instructions": None,
+            "instructions": (
+                "You are a helpful test assistant :)\n\n"
+                "Today is Friday 25/07/2025.\n\nAnswer in french."
+            ),
             "kind": "request",
             "parts": [
-                {
-                    "content": "You are a helpful test assistant :)",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Today is Friday 25/07/2025.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Answer in french.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
                 {
                     "content": ["Weather in Paris?"],
                     "part_kind": "user-prompt",
@@ -950,7 +902,10 @@ def test_post_conversation_tool_call_fails(api_client, mock_openai_stream_tool,
             "run_id": _run_id,
         },
         {
-            "instructions": None,
+            "instructions": (
+                "You are a helpful test assistant :)\n\n"
+                "Today is Friday 25/07/2025.\n\nAnswer in french."
+            ),
             "kind": "request",
             "parts": [
                 {
@@ -1214,27 +1169,11 @@ def test_post_conversation_data_protocol_no_stream(
     _run_id = chat_conversation.pydantic_messages[0]["run_id"]
     assert chat_conversation.pydantic_messages == [
         {
-            "instructions": None,
+            "instructions": (
+                "You are an amazing assistant.\n\nToday is Friday 25/07/2025.\n\nAnswer in english."
+            ),
             "kind": "request",
             "parts": [
-                {
-                    "content": "You are an amazing assistant.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Today is Friday 25/07/2025.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Answer in english.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
                 {
                     "content": ["Why the sky is blue?"],
                     "part_kind": "user-prompt",
@@ -1369,27 +1308,12 @@ async def test_post_conversation_async(api_client, mock_openai_stream, monkeypat
     _run_id = chat_conversation.pydantic_messages[0]["run_id"]
     assert chat_conversation.pydantic_messages == [
         {
-            "instructions": None,
+            "instructions": (
+                "You are a helpful test assistant :)\n\n"
+                "Today is Friday 25/07/2025.\n\nAnswer in english."
+            ),
             "kind": "request",
             "parts": [
-                {
-                    "content": "You are a helpful test assistant :)",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Today is Friday 25/07/2025.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
-                {
-                    "content": "Answer in english.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-07-25T10:36:35.297675Z",
-                },
                 {
                     "content": ["Hello"],
                     "part_kind": "user-prompt",

src/backend/chat/tests/views/chat/conversations/test_conversation_with_document_upload.py

@@ -216,7 +216,8 @@ def fixture_mock_openai_stream():
 @responses.activate
 @respx.mock
 @freeze_time()
-def test_post_conversation_with_document_upload(  # pylint: disable=too-many-arguments,too-many-positional-arguments
+def test_post_conversation_with_document_upload(
+    # pylint: disable=too-many-arguments,too-many-positional-arguments
     api_client,
     mock_albert_api,  # pylint: disable=unused-argument
     sample_pdf_content,
@@ -353,53 +354,25 @@ def test_post_conversation_with_document_upload(  # pylint: disable=too-many-arg
     assert len(chat_conversation.pydantic_messages) == 4
 
     _run_id = chat_conversation.pydantic_messages[0]["run_id"]
+
     assert chat_conversation.pydantic_messages[0] == {
-        "instructions": "When you receive a result from the summarization tool, you "
-        "MUST return it directly to the user without any "
-        "modification, paraphrasing, or additional summarization.The "
-        "tool already produces optimized summaries that should be "
-        "presented verbatim.You may translate the summary if "
-        "required, but you MUST preserve all the information from the "
-        "original summary.You may add a follow-up question after the "
-        "summary if needed.",
+        "instructions": "You are a helpful test assistant :)\n\n"
+        f"{today_promt_date}\n\n"
+        "Answer in english.\n\n"
+        "Use document_search_rag ONLY to retrieve specific passages from "
+        "attached documents. Do NOT use it to summarize; for summaries, "
+        "call the summarize tool instead.\n\nWhen you receive a result from the "
+        "summarization tool, you MUST return it directly to the user without "
+        "any modification, paraphrasing, or additional summarization."
+        "The tool already produces optimized summaries that should be "
+        "presented verbatim.You may translate the summary if required, "
+        "but you MUST preserve all the information from the original summary."
+        "You may add a follow-up question after the summary if needed.\n\n"
+        "[Internal context] User documents are attached to this conversation. "
+        "Do not request re-upload of documents; consider them already "
+        "available via the internal store.",
         "kind": "request",
         "parts": [
-            {
-                "content": "You are a helpful test assistant :)",
-                "dynamic_ref": None,
-                "part_kind": "system-prompt",
-                "timestamp": timezone_now,
-            },
-            {
-                "content": today_promt_date,
-                "dynamic_ref": None,
-                "part_kind": "system-prompt",
-                "timestamp": timezone_now,
-            },
-            {
-                "content": "Answer in english.",
-                "dynamic_ref": None,
-                "part_kind": "system-prompt",
-                "timestamp": timezone_now,
-            },
-            {
-                "content": "Use document_search_rag ONLY to retrieve specific "
-                "passages from attached documents. Do NOT use it to "
-                "summarize; for summaries, call the summarize tool "
-                "instead.",
-                "dynamic_ref": None,
-                "part_kind": "system-prompt",
-                "timestamp": timezone_now,
-            },
-            {
-                "content": "[Internal context] User documents are attached to this "
-                "conversation. Do not request re-upload of documents; "
-                "consider them already available via the internal "
-                "store.",
-                "dynamic_ref": None,
-                "part_kind": "system-prompt",
-                "timestamp": timezone_now,
-            },
             {
                 "content": ["What does the document say?"],
                 "part_kind": "user-prompt",
@@ -439,14 +412,21 @@ def test_post_conversation_with_document_upload(  # pylint: disable=too-many-arg
     }
     assert chat_conversation.pydantic_messages[2] == {
         "instructions": (
-            "When you receive a result from the summarization tool, you MUST "
-            "return it directly to the user without any modification, "
-            "paraphrasing, or additional summarization."
-            "The tool already produces optimized summaries that should "
-            "be presented verbatim."
-            "You may translate the summary if required, but you MUST preserve "
-            "all the information from the original summary."
-            "You may add a follow-up question after the summary if needed."
+            "You are a helpful test assistant :)\n\n"
+            f"{today_promt_date}\n\n"
+            "Answer in english.\n\n"
+            "Use document_search_rag ONLY to retrieve specific passages from "
+            "attached documents. Do NOT use it to summarize; for summaries, "
+            "call the summarize tool instead.\n\nWhen you receive a result from the "
+            "summarization tool, you MUST return it directly to the user without "
+            "any modification, paraphrasing, or additional summarization."
+            "The tool already produces optimized summaries that should be "
+            "presented verbatim.You may translate the summary if required, "
+            "but you MUST preserve all the information from the original summary."
+            "You may add a follow-up question after the summary if needed.\n\n"
+            "[Internal context] User documents are attached to this conversation. "
+            "Do not request re-upload of documents; consider them already "
+            "available via the internal store."
         ),
         "kind": "request",
         "parts": [
@@ -499,7 +479,8 @@ def test_post_conversation_with_document_upload(  # pylint: disable=too-many-arg
 @responses.activate
 @respx.mock
 @freeze_time("2025-07-25T10:36:35.297675Z")
-def test_post_conversation_with_document_upload_feature_disabled(  # pylint: disable=too-many-arguments,too-many-positional-arguments
+def test_post_conversation_with_document_upload_feature_disabled(
+    # pylint: disable=too-many-arguments,too-many-positional-arguments
     api_client,
     caplog,
     mock_openai_stream,  # pylint: disable=unused-argument
@@ -552,14 +533,12 @@ def test_post_conversation_with_document_upload_feature_disabled(  # pylint: dis
 
     # Replace UUIDs with placeholders for assertion
     response_content = replace_uuids_with_placeholder(response_content)
-
     assert response_content == (
         '0:"From the document, I can see that "\n'
         "0:\"it says 'Hello PDF'.\"\n"
         'f:{"messageId":"<mocked_uuid>"}\n'
         'd:{"finishReason":"stop","usage":{"promptTokens":150,"completionTokens":25}}\n'
     )
-
     # This behavior must be improved in the future to inform the user properly
     assert "Document upload feature is disabled, ignoring input documents." in caplog.text
 
@@ -582,6 +561,7 @@ def test_post_conversation_with_document_upload_summarize(  # pylint: disable=to
     api_client.force_authenticate(user=chat_conversation.owner)
 
     pdf_base64 = base64.b64encode(sample_pdf_content.read()).decode("utf-8")
+
     message = UIMessage(
         id="1",
         role="user",
@@ -643,7 +623,7 @@ def test_post_conversation_with_document_upload_summarize(  # pylint: disable=to
         'document discusses various topics."}\n'
         '0:"The document discusses various topics."\n'
         'f:{"messageId":"<mocked_uuid>"}\n'
-        'd:{"finishReason":"stop","usage":{"promptTokens":317,"completionTokens":19}}\n'
+        'd:{"finishReason":"stop","usage":{"promptTokens":287,"completionTokens":19}}\n'
     )
 
     # Check that the conversation was updated
@@ -705,52 +685,25 @@ def test_post_conversation_with_document_upload_summarize(  # pylint: disable=to
 
     _run_id = chat_conversation.pydantic_messages[0]["run_id"]
     assert chat_conversation.pydantic_messages[0] == {
-        "instructions": "When you receive a result from the summarization tool, you "
-        "MUST return it directly to the user without any "
-        "modification, paraphrasing, or additional summarization.The "
-        "tool already produces optimized summaries that should be "
-        "presented verbatim.You may translate the summary if "
-        "required, but you MUST preserve all the information from the "
-        "original summary.You may add a follow-up question after the "
-        "summary if needed.",
+        "instructions": (
+            "You are a helpful test assistant :)\n\n"
+            f"{today_promt_date}\n\n"
+            "Answer in english.\n\n"
+            "Use document_search_rag ONLY to retrieve specific passages from "
+            "attached documents. Do NOT use it to summarize; for summaries, "
+            "call the summarize tool instead.\n\nWhen you receive a result from the "
+            "summarization tool, you MUST return it directly to the user without "
+            "any modification, paraphrasing, or additional summarization."
+            "The tool already produces optimized summaries that should be "
+            "presented verbatim.You may translate the summary if required, "
+            "but you MUST preserve all the information from the original summary."
+            "You may add a follow-up question after the summary if needed.\n\n"
+            "[Internal context] User documents are attached to this conversation. "
+            "Do not request re-upload of documents; consider them already "
+            "available via the internal store."
+        ),
         "kind": "request",
         "parts": [
-            {
-                "content": "You are a helpful test assistant :)",
-                "dynamic_ref": None,
-                "part_kind": "system-prompt",
-                "timestamp": timezone_now,
-            },
-            {
-                "content": today_promt_date,
-                "dynamic_ref": None,
-                "part_kind": "system-prompt",
-                "timestamp": timezone_now,
-            },
-            {
-                "content": "Answer in english.",
-                "dynamic_ref": None,
-                "part_kind": "system-prompt",
-                "timestamp": timezone_now,
-            },
-            {
-                "content": "Use document_search_rag ONLY to retrieve specific "
-                "passages from attached documents. Do NOT use it to "
-                "summarize; for summaries, call the summarize tool "
-                "instead.",
-                "dynamic_ref": None,
-                "part_kind": "system-prompt",
-                "timestamp": timezone_now,
-            },
-            {
-                "content": "[Internal context] User documents are attached to this "
-                "conversation. Do not request re-upload of documents; "
-                "consider them already available via the internal "
-                "store.",
-                "dynamic_ref": None,
-                "part_kind": "system-prompt",
-                "timestamp": timezone_now,
-            },
             {
                 "content": ["Make a summary of this document."],
                 "part_kind": "user-prompt",
@@ -790,14 +743,21 @@ def test_post_conversation_with_document_upload_summarize(  # pylint: disable=to
     }
     assert chat_conversation.pydantic_messages[2] == {
         "instructions": (
-            "When you receive a result from the summarization tool, you MUST "
-            "return it directly to the user without any modification, "
-            "paraphrasing, or additional summarization."
-            "The tool already produces optimized summaries that should "
-            "be presented verbatim."
-            "You may translate the summary if required, but you MUST preserve "
-            "all the information from the original summary."
-            "You may add a follow-up question after the summary if needed."
+            "You are a helpful test assistant :)\n\n"
+            f"{today_promt_date}\n\n"
+            "Answer in english.\n\n"
+            "Use document_search_rag ONLY to retrieve specific passages from "
+            "attached documents. Do NOT use it to summarize; for summaries, "
+            "call the summarize tool instead.\n\nWhen you receive a result from the "
+            "summarization tool, you MUST return it directly to the user without "
+            "any modification, paraphrasing, or additional summarization."
+            "The tool already produces optimized summaries that should be "
+            "presented verbatim.You may translate the summary if required, "
+            "but you MUST preserve all the information from the original summary."
+            "You may add a follow-up question after the summary if needed.\n\n"
+            "[Internal context] User documents are attached to this conversation. "
+            "Do not request re-upload of documents; consider them already "
+            "available via the internal store."
         ),
         "kind": "request",
         "parts": [

src/backend/chat/tests/views/chat/conversations/test_conversation_with_document_url.py

@@ -17,7 +17,6 @@ from pydantic_ai.messages import (
     DocumentUrl,
     ModelMessage,
     ModelResponse,
-    SystemPromptPart,
     TextPart,
     UserPromptPart,
 )
@@ -61,7 +60,8 @@ def fixture_sample_document_content():
 
 @responses.activate
 @freeze_time()
-def test_post_conversation_with_local_pdf_document_url(  # pylint: disable=too-many-arguments,too-many-positional-arguments
+def test_post_conversation_with_local_pdf_document_url(
+    # pylint: disable=too-many-arguments,too-many-positional-arguments
     api_client,
     sample_document_content,
     today_promt_date,
@@ -120,7 +120,7 @@ def test_post_conversation_with_local_pdf_document_url(  # pylint: disable=too-m
     )
 
     async def agent_model(messages: list[ModelMessage], _info: AgentInfo):
-        presigned_url = messages[0].parts[3].content[1].url
+        presigned_url = messages[0].parts[0].content[1].url
         assert presigned_url.startswith("http://localhost:9000/conversations-media-storage/")
         assert presigned_url.find("X-Amz-Signature=") != -1
         assert presigned_url.find("X-Amz-Date=") != -1
@@ -129,11 +129,6 @@ def test_post_conversation_with_local_pdf_document_url(  # pylint: disable=too-m
         assert messages == [
             ModelRequest(
                 parts=[
-                    SystemPromptPart(
-                        content="You are a helpful test assistant :)", timestamp=timezone.now()
-                    ),
-                    SystemPromptPart(content=today_promt_date, timestamp=timezone.now()),
-                    SystemPromptPart(content="Answer in english.", timestamp=timezone.now()),
                     UserPromptPart(
                         content=[
                             "What is in this document?",
@@ -146,6 +141,8 @@ def test_post_conversation_with_local_pdf_document_url(  # pylint: disable=too-m
                         timestamp=timezone.now(),
                     ),
                 ],
+                instructions=f"You are a helpful test assistant :)\n\n{today_promt_date}"
+                "\n\nAnswer in english.",
                 run_id=messages[0].run_id,
             )
         ]
@@ -221,27 +218,11 @@ def test_post_conversation_with_local_pdf_document_url(  # pylint: disable=too-m
     _run_id = chat_conversation.pydantic_messages[0]["run_id"]
     assert chat_conversation.pydantic_messages == [
         {
-            "instructions": None,
+            "instructions": "You are a helpful test assistant :)\n\n"
+            f"{today_promt_date}\n\n"
+            "Answer in english.",
             "kind": "request",
             "parts": [
-                {
-                    "content": "You are a helpful test assistant :)",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": timestamp,
-                },
-                {
-                    "content": today_promt_date,
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": timestamp,
-                },
-                {
-                    "content": "Answer in english.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": timestamp,
-                },
                 {
                     "content": [
                         "What is in this document?",
@@ -429,7 +410,6 @@ def test_post_conversation_with_remote_document_url(
 @freeze_time("2025-10-18T20:48:20.286204Z")
 def test_post_conversation_with_local_document_url_in_history(  # pylint: disable=too-many-arguments,too-many-positional-arguments
     api_client,
-    today_promt_date,
     mock_ai_agent_service,
 ):
     """
@@ -437,6 +417,8 @@ def test_post_conversation_with_local_document_url_in_history(  # pylint: disabl
     """
     chat_conversation_pk = "0be55da5-8eb7-4dad-aa0f-fea454bd5809"
     document_url = f"/media-key/{chat_conversation_pk}/sample.pdf"
+    formatted_date = formats.date_format(timezone.now(), "l d/m/Y", use_l10n=False)
+
     chat_conversation = ChatConversationFactory(
         pk=chat_conversation_pk,
         owner__language="en-us",
@@ -472,27 +454,11 @@ def test_post_conversation_with_local_document_url_in_history(  # pylint: disabl
         ],
         pydantic_messages=[
             {
-                "instructions": None,
+                "instructions": "You are a helpful test assistant :)\n\n"
+                f"Today is {formatted_date}.\n\n"
+                "Answer in english.",
                 "kind": "request",
                 "parts": [
-                    {
-                        "content": "You are a helpful test assistant :)",
-                        "dynamic_ref": None,
-                        "part_kind": "system-prompt",
-                        "timestamp": "2025-10-18T20:48:20.286204Z",
-                    },
-                    {
-                        "content": today_promt_date,
-                        "dynamic_ref": None,
-                        "part_kind": "system-prompt",
-                        "timestamp": "2025-10-18T20:48:20.286204Z",
-                    },
-                    {
-                        "content": "Answer in english.",
-                        "dynamic_ref": None,
-                        "part_kind": "system-prompt",
-                        "timestamp": "2025-10-18T20:48:20.286204Z",
-                    },
                     {
                         "content": [
                             "What is in this document?",
@@ -555,7 +521,7 @@ def test_post_conversation_with_local_document_url_in_history(  # pylint: disabl
     )
 
     async def agent_model(messages: list[ModelMessage], _info: AgentInfo):
-        presigned_url = messages[0].parts[3].content[1].url
+        presigned_url = messages[0].parts[0].content[1].url
         assert presigned_url.startswith("http://localhost:9000/conversations-media-storage/")
         assert presigned_url.find("X-Amz-Signature=") != -1
         assert presigned_url.find("X-Amz-Date=") != -1
@@ -564,18 +530,6 @@ def test_post_conversation_with_local_document_url_in_history(  # pylint: disabl
         assert messages == [
             ModelRequest(
                 parts=[
-                    SystemPromptPart(
-                        content="You are a helpful test assistant :)",
-                        timestamp=timezone.now(),
-                    ),
-                    SystemPromptPart(
-                        content=today_promt_date,
-                        timestamp=timezone.now(),
-                    ),
-                    SystemPromptPart(
-                        content="Answer in english.",
-                        timestamp=timezone.now(),
-                    ),
                     UserPromptPart(
                         content=[
                             "What is in this document?",
@@ -588,6 +542,9 @@ def test_post_conversation_with_local_document_url_in_history(  # pylint: disabl
                         timestamp=timezone.now(),
                     ),
                 ],
+                instructions="You are a helpful test assistant :)\n\n"
+                "Today is Saturday 18/10/2025.\n\n"
+                "Answer in english.",
                 run_id=messages[0].run_id,
             ),
             ModelResponse(
@@ -606,6 +563,9 @@ def test_post_conversation_with_local_document_url_in_history(  # pylint: disabl
                         timestamp=timezone.now(),
                     )
                 ],
+                instructions="You are a helpful test assistant :)\n\n"
+                "Today is Saturday 18/10/2025.\n\n"
+                "Answer in english.",
                 run_id=messages[2].run_id,
             ),
         ]
@@ -705,27 +665,11 @@ def test_post_conversation_with_local_document_url_in_history(  # pylint: disabl
     _run_id = chat_conversation.pydantic_messages[2]["run_id"]
     assert chat_conversation.pydantic_messages == [
         {
-            "instructions": None,
+            "instructions": "You are a helpful test assistant :)\n\n"
+            "Today is Saturday 18/10/2025.\n\n"
+            "Answer in english.",
             "kind": "request",
             "parts": [
-                {
-                    "content": "You are a helpful test assistant :)",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-10-18T20:48:20.286204Z",
-                },
-                {
-                    "content": today_promt_date,
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-10-18T20:48:20.286204Z",
-                },
-                {
-                    "content": "Answer in english.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-10-18T20:48:20.286204Z",
-                },
                 {
                     "content": [
                         "What is in this document?",
@@ -772,7 +716,9 @@ def test_post_conversation_with_local_document_url_in_history(  # pylint: disabl
             # no run_id here
         },
         {
-            "instructions": None,
+            "instructions": "You are a helpful test assistant :)\n\n"
+            "Today is Saturday 18/10/2025.\n\n"
+            "Answer in english.",
             "kind": "request",
             "parts": [
                 {
@@ -823,7 +769,8 @@ def test_post_conversation_with_local_document_url_in_history(  # pylint: disabl
         ("data.csv", "text/csv"),
     ],
 )
-def test_post_conversation_with_local_not_pdf_document_url(  # pylint: disable=too-many-arguments,too-many-positional-arguments
+def test_post_conversation_with_local_not_pdf_document_url(
+    # pylint: disable=too-many-arguments,too-many-positional-arguments
     api_client,
     today_promt_date,
     mock_ai_agent_service,
@@ -886,27 +833,6 @@ def test_post_conversation_with_local_not_pdf_document_url(  # pylint: disable=t
         assert messages == [
             ModelRequest(
                 parts=[
-                    SystemPromptPart(
-                        content="You are a helpful test assistant :)", timestamp=timezone.now()
-                    ),
-                    SystemPromptPart(content=today_promt_date, timestamp=timezone.now()),
-                    SystemPromptPart(content="Answer in english.", timestamp=timezone.now()),
-                    SystemPromptPart(
-                        content=(
-                            "Use document_search_rag ONLY to retrieve specific passages from "
-                            "attached documents. Do NOT use it to summarize; for summaries, "
-                            "call the summarize tool instead."
-                        ),
-                        timestamp=timezone.now(),
-                    ),
-                    SystemPromptPart(
-                        content=(
-                            "[Internal context] User documents are attached to this conversation. "
-                            "Do not request re-upload of documents; consider them already "
-                            "available via the internal store."
-                        ),
-                        timestamp=timezone.now(),
-                    ),
                     UserPromptPart(
                         content=[
                             "What is in this document?",
@@ -916,14 +842,22 @@ def test_post_conversation_with_local_not_pdf_document_url(  # pylint: disable=t
                     ),
                 ],
                 instructions=(
-                    "When you receive a result from the summarization tool, you MUST "
-                    "return it directly to the user without any modification, "
-                    "paraphrasing, or additional summarization."
-                    "The tool already produces optimized summaries that should "
-                    "be presented verbatim."
-                    "You may translate the summary if required, but you MUST preserve "
-                    "all the information from the original summary."
-                    "You may add a follow-up question after the summary if needed."
+                    "You are a helpful test assistant :)\n\n"
+                    f"{today_promt_date}\n\n"
+                    "Answer in english.\n\n"
+                    "Use document_search_rag ONLY to retrieve specific passages from "
+                    "attached documents. Do NOT use it to summarize; for summaries, "
+                    "call the summarize tool instead.\n\nWhen you receive a result "
+                    "from the summarization tool, you MUST return it directly to "
+                    "the user without any modification, paraphrasing, or additional "
+                    "summarization.The tool already produces optimized summaries "
+                    "that should be presented verbatim.You may translate the summary "
+                    "if required, but you MUST preserve all the information from the "
+                    "original summary.You may add a follow-up question after the "
+                    "summary if needed.\n\n"
+                    "[Internal context] User documents are attached to this conversation. "
+                    "Do not request re-upload of documents; "
+                    "consider them already available via the internal store."
                 ),
                 run_id=messages[0].run_id,
             )
@@ -999,53 +933,25 @@ def test_post_conversation_with_local_not_pdf_document_url(  # pylint: disable=t
     assert chat_conversation.pydantic_messages == [
         {
             "instructions": (
-                "When you receive a result from the summarization tool, you MUST "
-                "return it directly to the user without any modification, "
-                "paraphrasing, or additional summarization."
-                "The tool already produces optimized summaries that should "
-                "be presented verbatim."
-                "You may translate the summary if required, but you MUST preserve "
-                "all the information from the original summary."
-                "You may add a follow-up question after the summary if needed."
+                "You are a helpful test assistant :)\n\n"
+                f"{today_promt_date}\n\n"
+                "Answer in english.\n\n"
+                "Use document_search_rag ONLY to retrieve specific passages from "
+                "attached documents. Do NOT use it to summarize; for summaries, "
+                "call the summarize tool instead.\n\nWhen you receive a result "
+                "from the summarization tool, you MUST return it directly to "
+                "the user without any modification, paraphrasing, or additional "
+                "summarization.The tool already produces optimized summaries "
+                "that should be presented verbatim.You may translate the summary "
+                "if required, but you MUST preserve all the information from the "
+                "original summary.You may add a follow-up question after the "
+                "summary if needed.\n\n"
+                "[Internal context] User documents are attached to this conversation. "
+                "Do not request re-upload of documents; "
+                "consider them already available via the internal store."
             ),
             "kind": "request",
             "parts": [
-                {
-                    "content": "You are a helpful test assistant :)",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": timestamp,
-                },
-                {
-                    "content": today_promt_date,
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": timestamp,
-                },
-                {
-                    "content": "Answer in english.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": timestamp,
-                },
-                {
-                    "content": "Use document_search_rag ONLY to retrieve specific "
-                    "passages from attached documents. Do NOT use it to "
-                    "summarize; for summaries, call the summarize tool "
-                    "instead.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": timestamp,
-                },
-                {
-                    "content": "[Internal context] User documents are attached to "
-                    "this conversation. Do not request re-upload of "
-                    "documents; consider them already available via the "
-                    "internal store.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": timestamp,
-                },
                 {
                     "content": [
                         "What is in this document?",

src/backend/chat/tests/views/chat/conversations/test_conversation_with_history.py

@@ -919,7 +919,7 @@ def history_conversation_with_tool_fixture():
     history_timestamp = timezone.now().replace(year=2025, month=6, day=15, hour=10, minute=30)
 
     # Create a conversation with pre-existing messages including a tool invocation
-    conversation = ChatConversationFactory()
+    conversation = ChatConversationFactory(owner__language="nl-nl")
 
     # Add previous user and assistant messages with tool invocation
     conversation.messages = [
@@ -1377,7 +1377,9 @@ def test_post_conversation_with_existing_tool_history(
 
     # Verify the new tool call request is included
     assert history_conversation_with_tool.pydantic_messages[8] == {
-        "instructions": None,
+        "instructions": "You are a helpful test assistant :)\n\n"
+        "Today is Friday 25/07/2025.\n\n"
+        "Answer in dutch.",
         "kind": "request",
         "parts": [
             {
@@ -1420,7 +1422,9 @@ def test_post_conversation_with_existing_tool_history(
     }
 
     assert history_conversation_with_tool.pydantic_messages[10] == {
-        "instructions": None,
+        "instructions": "You are a helpful test assistant :)\n\n"
+        "Today is Friday 25/07/2025.\n\n"
+        "Answer in dutch.",
         "kind": "request",
         "parts": [
             {

src/backend/chat/tests/views/chat/conversations/test_conversation_with_image_url.py

@@ -2,7 +2,7 @@
 
 import uuid
 
-from django.utils import timezone
+from django.utils import formats, timezone
 
 import pytest
 from dirty_equals import IsUUID
@@ -12,7 +12,6 @@ from pydantic_ai.messages import (
     ImageUrl,
     ModelMessage,
     ModelResponse,
-    SystemPromptPart,
     TextPart,
     UserPromptPart,
 )
@@ -87,22 +86,15 @@ def test_post_conversation_with_local_image_url(
     )
 
     async def agent_model(messages: list[ModelMessage], _info: AgentInfo):
-        presigned_url = messages[0].parts[3].content[1].url
-        assert presigned_url.startswith("http://localhost:9000/conversations-media-storage/")
+        presigned_url = messages[0].parts[0].content[1].url
+        # assert presigned_url.startswith("http://localhost:9000/conversations-media-storage/")
         assert presigned_url.find("X-Amz-Signature=") != -1
         assert presigned_url.find("X-Amz-Date=") != -1
         assert presigned_url.find("X-Amz-Expires=") != -1
-
+        formatted_date = formats.date_format(timezone.now(), "l d/m/Y", use_l10n=False)
         assert messages == [
             ModelRequest(
                 parts=[
-                    SystemPromptPart(
-                        content="You are a helpful test assistant :)", timestamp=timezone.now()
-                    ),
-                    SystemPromptPart(
-                        content="Today is Saturday 18/10/2025.", timestamp=timezone.now()
-                    ),
-                    SystemPromptPart(content="Answer in english.", timestamp=timezone.now()),
                     UserPromptPart(
                         content=[
                             "What is in this image?",
@@ -115,6 +107,8 @@ def test_post_conversation_with_local_image_url(
                         timestamp=timezone.now(),
                     ),
                 ],
+                instructions="You are a helpful test assistant :)\n\nToday is "
+                f"{formatted_date}.\n\nAnswer in english.",
                 run_id=messages[0].run_id,
             )
         ]
@@ -184,27 +178,10 @@ def test_post_conversation_with_local_image_url(
     _run_id = chat_conversation.pydantic_messages[0]["run_id"]
     assert chat_conversation.pydantic_messages == [
         {
-            "instructions": None,
+            "instructions": "You are a helpful test assistant :)\n\n"
+            "Today is Saturday 18/10/2025.\n\nAnswer in english.",
             "kind": "request",
             "parts": [
-                {
-                    "content": "You are a helpful test assistant :)",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-10-18T20:48:20.286204Z",
-                },
-                {
-                    "content": "Today is Saturday 18/10/2025.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-10-18T20:48:20.286204Z",
-                },
-                {
-                    "content": "Answer in english.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-10-18T20:48:20.286204Z",
-                },
                 {
                     "content": [
                         "What is in this image?",
@@ -286,11 +263,6 @@ def test_post_conversation_with_local_image_wrong_url(
         assert messages == [
             ModelRequest(
                 parts=[
-                    SystemPromptPart(
-                        content="You are a helpful test assistant :)", timestamp=timezone.now()
-                    ),
-                    SystemPromptPart(content=today_promt_date, timestamp=timezone.now()),
-                    SystemPromptPart(content="Answer in english.", timestamp=timezone.now()),
                     UserPromptPart(
                         content=[
                             "What is in this image?",
@@ -303,6 +275,8 @@ def test_post_conversation_with_local_image_wrong_url(
                         timestamp=timezone.now(),
                     ),
                 ],
+                instructions=f"You are a helpful test assistant :)\n\n{today_promt_date}"
+                "\n\nAnswer in english.",
                 run_id=messages[0].run_id,
             )
         ]
@@ -374,11 +348,6 @@ def test_post_conversation_with_remote_image_url(
         assert messages == [
             ModelRequest(
                 parts=[
-                    SystemPromptPart(
-                        content="You are a helpful test assistant :)", timestamp=timezone.now()
-                    ),
-                    SystemPromptPart(content=today_promt_date, timestamp=timezone.now()),
-                    SystemPromptPart(content="Answer in english.", timestamp=timezone.now()),
                     UserPromptPart(
                         content=[
                             "What is in this image?",
@@ -391,6 +360,8 @@ def test_post_conversation_with_remote_image_url(
                         timestamp=timezone.now(),
                     ),
                 ],
+                instructions="You are a helpful test assistant :)\n\n"
+                f"{today_promt_date}\n\nAnswer in english.",
                 run_id=messages[0].run_id,
             )
         ]
@@ -504,27 +475,10 @@ def test_post_conversation_with_local_image_url_in_history(
         ],
         pydantic_messages=[
             {
-                "instructions": None,
+                "instructions": f"You are a helpful test assistant :)\n\n{today_promt_date}"
+                "\n\nAnswer in english.",
                 "kind": "request",
                 "parts": [
-                    {
-                        "content": "You are a helpful test assistant :)",
-                        "dynamic_ref": None,
-                        "part_kind": "system-prompt",
-                        "timestamp": "2025-10-18T20:48:20.286204Z",
-                    },
-                    {
-                        "content": today_promt_date,
-                        "dynamic_ref": None,
-                        "part_kind": "system-prompt",
-                        "timestamp": "2025-10-18T20:48:20.286204Z",
-                    },
-                    {
-                        "content": "Answer in english.",
-                        "dynamic_ref": None,
-                        "part_kind": "system-prompt",
-                        "timestamp": "2025-10-18T20:48:20.286204Z",
-                    },
                     {
                         "content": [
                             "What is in this image?",
@@ -587,7 +541,7 @@ def test_post_conversation_with_local_image_url_in_history(
     )
 
     async def agent_model(messages: list[ModelMessage], _info: AgentInfo):
-        presigned_url = messages[0].parts[3].content[1].url
+        presigned_url = messages[0].parts[0].content[1].url
         assert presigned_url.startswith("http://localhost:9000/conversations-media-storage/")
         assert presigned_url.find("X-Amz-Signature=") != -1
         assert presigned_url.find("X-Amz-Date=") != -1
@@ -596,18 +550,6 @@ def test_post_conversation_with_local_image_url_in_history(
         assert messages == [
             ModelRequest(
                 parts=[
-                    SystemPromptPart(
-                        content="You are a helpful test assistant :)",
-                        timestamp=timezone.now(),
-                    ),
-                    SystemPromptPart(
-                        content=today_promt_date,
-                        timestamp=timezone.now(),
-                    ),
-                    SystemPromptPart(
-                        content="Answer in english.",
-                        timestamp=timezone.now(),
-                    ),
                     UserPromptPart(
                         content=[
                             "What is in this image?",
@@ -619,7 +561,9 @@ def test_post_conversation_with_local_image_url_in_history(
                         ],
                         timestamp=timezone.now(),
                     ),
-                ]
+                ],
+                instructions="You are a helpful test assistant :)\n\n"
+                f"{today_promt_date}\n\nAnswer in english.",
             ),
             ModelResponse(
                 parts=[TextPart(content="This is an image of a single pixel.")],
@@ -637,6 +581,8 @@ def test_post_conversation_with_local_image_url_in_history(
                     )
                 ],
                 run_id=messages[2].run_id,
+                instructions="You are a helpful test assistant :)\n\n"
+                "Today is Saturday 18/10/2025.\n\nAnswer in english.",
             ),
         ]
         yield "This is an image of square, very small and nice."
@@ -735,27 +681,10 @@ def test_post_conversation_with_local_image_url_in_history(
     _run_id = chat_conversation.pydantic_messages[2]["run_id"]
     assert chat_conversation.pydantic_messages == [
         {
-            "instructions": None,
+            "instructions": f"You are a helpful test assistant :)\n\n{today_promt_date}"
+            "\n\nAnswer in english.",
             "kind": "request",
             "parts": [
-                {
-                    "content": "You are a helpful test assistant :)",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-10-18T20:48:20.286204Z",
-                },
-                {
-                    "content": today_promt_date,
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-10-18T20:48:20.286204Z",
-                },
-                {
-                    "content": "Answer in english.",
-                    "dynamic_ref": None,
-                    "part_kind": "system-prompt",
-                    "timestamp": "2025-10-18T20:48:20.286204Z",
-                },
                 {
                     "content": [
                         "What is in this image?",
@@ -796,7 +725,8 @@ def test_post_conversation_with_local_image_url_in_history(
             },
         },
         {
-            "instructions": None,
+            "instructions": "You are a helpful test assistant :)\n\nToday is Saturday 18/10/2025."
+            "\n\nAnswer in english.",
             "kind": "request",
             "parts": [
                 {

src/backend/chat/tests/views/chat/conversations/test_stop_streaming.py

@@ -1,4 +1,4 @@
-"""Test the post_stop_steaming view."""
+"""Test the post_stop_streaming view."""
 
 from unittest.mock import patch
 

src/backend/chat/tools/data_analysis.py

@@ -0,0 +1,873 @@
+"""Data analysis tool for tabular files (CSV, Excel)."""
+
+import base64
+import functools
+import json
+import logging
+import uuid
+from io import BytesIO
+from typing import Any, Dict
+
+import matplotlib
+import numpy as np
+
+matplotlib.use("Agg")  # Use non-interactive backend
+import matplotlib.pyplot as plt
+import pandas as pd
+
+from django.conf import settings
+from django.core.files.storage import default_storage
+from django.db.models import Q
+from asgiref.sync import sync_to_async
+from pydantic_ai import Agent, RunContext
+from pydantic_ai.exceptions import ModelRetry
+from pydantic_ai.messages import ToolReturn
+
+from core.file_upload.enums import AttachmentStatus
+from core.file_upload.utils import generate_retrieve_policy
+
+from chat.agents.base import BaseAgent, prepare_custom_model
+from chat.models import ChatConversationAttachment
+from chat.tools.exceptions import ModelCannotRetry
+from chat.tools.utils import last_model_retry_soft_fail
+
+logger = logging.getLogger(__name__)
+
+# MIME types for tabular files
+TABULAR_MIME_TYPES = [
+    "text/csv",
+    "application/csv",
+    "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
+    "application/vnd.ms-excel",
+    "application/excel",
+]
+
+
+@sync_to_async
+def read_tabular_file(attachment) -> bytes:
+    """Read tabular file content asynchronously."""
+    with default_storage.open(attachment.key, "rb") as f:
+        return f.read()
+
+
+def detect_csv_separator(file_data: bytes) -> str:
+    """
+    Detect the CSV separator by analyzing the first few lines.
+    
+    Returns the most likely separator: ',', ';', or '\t'
+    """
+    # Read first 10KB to analyze
+    sample = file_data[:10240].decode("utf-8", errors="ignore")
+    lines = sample.split("\n")[:10]  # First 10 lines
+    
+    if not lines:
+        return ","  # Default to comma
+    
+    # Count occurrences of each separator in the first few lines
+    comma_count = sum(line.count(",") for line in lines)
+    semicolon_count = sum(line.count(";") for line in lines)
+    tab_count = sum(line.count("\t") for line in lines)
+    
+    # Return the separator with the highest count
+    if tab_count > comma_count and tab_count > semicolon_count:
+        return "\t"
+    elif semicolon_count > comma_count:
+        return ";"
+    else:
+        return ","  # Default to comma
+
+
+def _convert_to_serializable(obj: Any) -> Any:
+    """
+    Convert pandas/numpy types to Python native types for JSON serialization.
+    
+    Handles:
+    - pandas DataFrame/Series
+    - numpy scalars (int64, float64, etc.)
+    - numpy arrays
+    - pandas Timestamp
+    - Other non-serializable types
+    
+    Args:
+        obj: The object to convert.
+        
+    Returns:
+        A JSON-serializable version of the object.
+    """
+    
+    # Handle pandas DataFrame
+    if isinstance(obj, pd.DataFrame):
+        # Limit number of rows to avoid huge responses
+        if len(obj) > 1000:
+            obj = obj.head(1000)
+            logger.warning("Result truncated to 1000 rows")
+        return obj.to_dict(orient="records")
+    
+    # Handle pandas Series
+    if isinstance(obj, pd.Series):
+        # Convert Series to dict, handling index
+        result_dict = obj.to_dict()
+        # Convert any numpy/pandas types in the values
+        return {str(k): _convert_to_serializable(v) for k, v in result_dict.items()}
+    
+    # Handle numpy scalars
+    if isinstance(obj, (np.integer, np.floating)):
+        return obj.item()  # Convert to Python native int/float
+    
+    # Handle numpy arrays
+    if isinstance(obj, np.ndarray):
+        return obj.tolist()
+    
+    # Handle pandas Timestamp
+    if isinstance(obj, pd.Timestamp):
+        return obj.isoformat()
+    
+    # Handle lists and tuples - recursively convert elements
+    if isinstance(obj, (list, tuple)):
+        return [_convert_to_serializable(item) for item in obj]
+    
+    # Handle dicts - recursively convert values
+    if isinstance(obj, dict):
+        return {str(k): _convert_to_serializable(v) for k, v in obj.items()}
+    
+    # Handle None, bool, int, float, str - these are already serializable
+    if obj is None or isinstance(obj, (bool, int, float, str)):
+        return obj
+    
+    # Fallback: try to convert to string
+    try:
+        return str(obj)
+    except Exception:
+        logger.warning("Could not serialize object of type %s, returning None", type(obj))
+        return None
+
+
+def _is_valid_excel_file(file_data: bytes, file_name: str) -> bool:
+    """
+    Check if the file data appears to be a valid Excel file.
+    
+    XLSX files are ZIP archives, so they should start with ZIP signature (PK\x03\x04).
+    XLS files have a different signature.
+    """
+    if not file_data:
+        return False
+    
+    file_lower = file_name.lower()
+    
+    # Check for XLSX (ZIP-based) signature
+    if file_lower.endswith((".xlsx", ".xlsm", ".xlsb")):
+        # XLSX files are ZIP archives, should start with PK\x03\x04
+        return file_data[:4] == b"PK\x03\x04"
+    
+    # Check for XLS (OLE2) signature
+    if file_lower.endswith(".xls"):
+        # XLS files are OLE2 compound documents, should start with specific signature
+        # Common signatures: 0xD0CF11E0 (OLE2) or 0x504B0304 (sometimes saved as ZIP)
+        return (
+            file_data[:4] == b"\xd0\xcf\x11\xe0"  # OLE2 signature
+            or file_data[:4] == b"PK\x03\x04"  # Sometimes XLS are actually ZIP
+        )
+    
+    return False
+
+
+@sync_to_async
+def load_dataframe(file_data: bytes, content_type: str, file_name: str) -> Dict[str, pd.DataFrame]:
+    """
+    Load tabular file into pandas DataFrames.
+    
+    Returns a dictionary mapping sheet/table names to DataFrames.
+    For CSV files, uses 'default' as the key.
+    For Excel files, uses sheet names as keys.
+    """
+    try:
+        # Handle CSV files - also accept text/plain if file extension is .csv
+        if content_type in ["text/csv", "application/csv"] or (
+            content_type == "text/plain" and file_name.lower().endswith(".csv")
+        ):
+            # Detect the separator
+            separator = detect_csv_separator(file_data)
+            logger.debug("Detected CSV separator: %r", separator)
+            
+            # Read CSV with detected separator
+            df = pd.read_csv(
+                BytesIO(file_data),
+                sep=separator,
+                on_bad_lines="skip",  # Skip problematic lines
+                engine="python",  # More flexible parser
+                encoding="utf-8",
+            )
+            
+            if df.empty:
+                raise ValueError("CSV file appears to be empty or could not be parsed")
+            
+            return {"default": df}
+        elif content_type in [
+            "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
+            "application/vnd.ms-excel",
+            "application/excel",
+        ] or file_name.lower().endswith((".xlsx", ".xls", ".xlsm", ".xlsb")):
+            # Validate Excel file format before attempting to read
+            if not _is_valid_excel_file(file_data, file_name):
+                logger.warning(
+                    "File '%s' does not appear to be a valid Excel file. "
+                    "File size: %d bytes, First bytes: %r",
+                    file_name,
+                    len(file_data),
+                    file_data[:20] if len(file_data) >= 20 else file_data,
+                )
+                raise ValueError(
+                    f"File '{file_name}' does not appear to be a valid Excel file. "
+                    "It may be corrupted or in an unsupported format."
+                )
+            
+            file_lower = file_name.lower()
+            dataframes = {}
+            
+            # Try different engines based on file extension
+            if file_lower.endswith(".xls"):
+                # Old Excel format - try xlrd engine
+                try:
+                    logger.debug("Attempting to read .xls file with xlrd engine")
+                    excel_file = pd.ExcelFile(BytesIO(file_data), engine="xlrd")
+                    dataframes = {
+                        sheet_name: excel_file.parse(sheet_name)
+                        for sheet_name in excel_file.sheet_names
+                    }
+                except Exception as xlrd_error:
+                    logger.warning("Failed to read .xls with xlrd: %s", xlrd_error)
+                    # Fallback: try openpyxl (sometimes .xls files are actually .xlsx)
+                    try:
+                        logger.debug("Trying openpyxl as fallback for .xls file")
+                        excel_file = pd.ExcelFile(BytesIO(file_data), engine="openpyxl")
+                        dataframes = {
+                            sheet_name: excel_file.parse(sheet_name)
+                            for sheet_name in excel_file.sheet_names
+                        }
+                    except Exception as openpyxl_error:
+                        logger.error("Failed to read .xls with both engines: %s", openpyxl_error)
+                        raise ValueError(
+                            f"Failed to read Excel file '{file_name}': "
+                            f"xlrd error: {xlrd_error}, openpyxl error: {openpyxl_error}"
+                        ) from openpyxl_error
+            else:
+                # XLSX format - try openpyxl first
+                try:
+                    logger.debug("Attempting to read Excel file with openpyxl engine")
+                    excel_file = pd.ExcelFile(BytesIO(file_data), engine="openpyxl")
+                    dataframes = {
+                        sheet_name: excel_file.parse(sheet_name)
+                        for sheet_name in excel_file.sheet_names
+                    }
+                except Exception as openpyxl_error:
+                    logger.warning("Failed to read with openpyxl: %s", openpyxl_error)
+                    # Try calamine engine if available (faster and more robust)
+                    try:
+                        logger.debug("Trying calamine engine as fallback")
+                        excel_file = pd.ExcelFile(BytesIO(file_data), engine="calamine")
+                        dataframes = {
+                            sheet_name: excel_file.parse(sheet_name)
+                            for sheet_name in excel_file.sheet_names
+                        }
+                    except ImportError:
+                        logger.debug("calamine engine not available")
+                        raise ValueError(
+                            f"Failed to read Excel file '{file_name}' with openpyxl: {openpyxl_error}. "
+                            "The file may be corrupted or in an unsupported format."
+                        ) from openpyxl_error
+                    except Exception as calamine_error:
+                        logger.error("Failed to read with calamine: %s", calamine_error)
+                        raise ValueError(
+                            f"Failed to read Excel file '{file_name}': "
+                            f"openpyxl error: {openpyxl_error}, calamine error: {calamine_error}"
+                        ) from calamine_error
+            
+            if not dataframes:
+                raise ValueError(f"Excel file '{file_name}' contains no readable sheets")
+            
+            logger.info(
+                "Successfully loaded Excel file '%s' with %d sheet(s): %s",
+                file_name,
+                len(dataframes),
+                list(dataframes.keys()),
+            )
+            return dataframes
+        else:
+            raise ValueError(f"Unsupported content type: {content_type}")
+    except Exception as e:
+        logger.error("Error loading tabular file: %s", e, exc_info=True)
+        raise ModelCannotRetry(f"Failed to load file: {str(e)}") from e
+
+
+def generate_metadata(dataframes: Dict[str, pd.DataFrame], file_name: str) -> Dict[str, Any]:
+    """
+    Generate metadata about the tabular file.
+    
+    Returns:
+        Dictionary containing:
+        - sheets: List of sheet/table names
+        - schemas: Dictionary mapping sheet names to their schemas
+        - row_counts: Dictionary mapping sheet names to row counts
+        - column_info: Dictionary mapping sheet names to column information
+    """
+    metadata = {
+        "file_name": file_name,
+        "sheets": list(dataframes.keys()),
+        "schemas": {},
+        "row_counts": {},
+        "column_info": {},
+    }
+
+    for sheet_name, df in dataframes.items():
+        # Schema: column names and types
+        metadata["schemas"][sheet_name] = {
+            col: str(dtype) for col, dtype in df.dtypes.items()
+        }
+        
+        # Row count
+        metadata["row_counts"][sheet_name] = len(df)
+        
+        # Column info: name, type, sample values, null counts
+        metadata["column_info"][sheet_name] = {}
+        for col in df.columns:
+            col_info = {
+                "type": str(df[col].dtype),
+                "null_count": int(df[col].isna().sum()),
+                "unique_count": int(df[col].nunique()),
+            }
+            # Add sample values (non-null)
+            sample_values = df[col].dropna().head(5).tolist()
+            if sample_values:
+                col_info["sample_values"] = [str(v) for v in sample_values]
+            metadata["column_info"][sheet_name][col] = col_info
+
+    return metadata
+
+
+async def generate_query(
+    user_query: str, metadata: Dict[str, Any], query_agent: BaseAgent, ctx: RunContext
+) -> str:
+    """
+    Use an LLM agent to generate a pandas query from user query and file metadata.
+    """
+    metadata_str = json.dumps(metadata, indent=2)
+    
+    prompt = f"""You are a data analysis assistant. Given a user query and file metadata, generate a Python pandas query to answer the question.
+
+File metadata:
+{metadata_str}
+
+User query: {user_query}
+
+Generate a Python code snippet that:
+1. Uses pandas operations (filter, groupby, aggregate, etc.)
+2. Works with the dataframes loaded in memory (available as 'dataframes' dict)
+3. Assigns the final result to a variable named 'result'
+4. Handles the specific sheet/table if multiple sheets exist
+5. ALWAYS handles NaN/NA values in boolean conditions using .notna() or .fillna() before filtering
+6. If the user asks for a plot/graph/chart, create it using matplotlib and save to 'plot_image' variable as base64
+
+IMPORTANT RULES:
+- The code MUST assign the final result to a variable named 'result'
+- When filtering with conditions, ALWAYS check for NaN first: df[df['col'].notna() & (df['col'] > value)]
+- Use .dropna() if you need to remove rows with missing values
+- Use .fillna() if you need to replace missing values
+- If plotting: use plt (already imported), create the plot, convert to base64:
+  ```python
+  plt.figure(figsize=(10, 6))
+  # ... your plot code ...
+  buf = BytesIO()
+  plt.savefig(buf, format='png')
+  buf.seek(0)
+  plot_image = base64.b64encode(buf.getvalue()).decode('utf-8')
+  plt.close()
+  ```
+  NOTE: Do NOT use import statements - plt, base64, BytesIO are already available.
+
+Return ONLY the Python code, without markdown formatting or explanations. The code should be executable and use variables:
+- 'dataframes': dict mapping sheet names to DataFrames
+- Sheet names available: {', '.join(metadata['sheets'])}
+
+Example format (without plot):
+df = dataframes['default']
+df = df[df['column'].notna()]  # Remove NaN values first
+result = df[df['column'] > 100].groupby('category').sum()
+
+Example format (with plot):
+df = dataframes['default']
+plt.figure(figsize=(10, 6))
+plt.plot(df.index, df['close'])
+plt.xlabel('Index')
+plt.ylabel('Close')
+plt.title('Close vs Index')
+buf = BytesIO()
+plt.savefig(buf, format='png')
+buf.seek(0)
+plot_image = base64.b64encode(buf.getvalue()).decode('utf-8')
+plt.close()
+result = "Plot generated successfully. The plot image has been saved and is available in the tool response."
+
+IMPORTANT: 
+- Do NOT use import statements in the code. All necessary modules (pd, plt, np, base64, BytesIO) are already available. Do NOT use anything else than these modules.
+- When returning the result text, mention that a plot was generated and will be available in the response, but do NOT include the URL in the text - the system will handle displaying it.
+
+Generate the query code:"""
+
+    try:
+        response = await query_agent.run(prompt, usage=ctx.usage)
+        query_code = response.output.strip()
+        
+        # Extract code from markdown code blocks if present
+        if "```python" in query_code:
+            query_code = query_code.split("```python")[1].split("```")[0].strip()
+        elif "```" in query_code:
+            query_code = query_code.split("```")[1].split("```")[0].strip()
+        
+        return query_code
+    except Exception as e:
+        logger.error("Error generating query: %s", e, exc_info=True)
+        raise ModelRetry("Failed to generate query. Please try rephrasing your question.") from e
+
+
+@sync_to_async
+def execute_query(query_code: str, dataframes: Dict[str, pd.DataFrame]) -> Any:
+    """
+    Execute the generated pandas query safely.
+    
+    Note: Uses exec() in a restricted environment. The query code is generated
+    by an LLM based on file metadata, so it should be relatively safe, but
+    we restrict the available builtins and globals.
+    """
+    try:
+        # Pre-process dataframes to handle common issues
+        processed_dataframes = {}
+        for name, df in dataframes.items():
+            # Make a copy to avoid modifying original
+            df_processed = df.copy()
+            # Replace common NaN representations
+            df_processed = df_processed.replace(["", " ", "nan", "NaN", "None", "null"], pd.NA)
+            processed_dataframes[name] = df_processed
+        
+        # Create a safe execution environment
+        safe_globals = {
+            "pd": pd,
+            "plt": plt,
+            "np": np,
+            "base64": base64,
+            "BytesIO": BytesIO,
+            "dataframes": processed_dataframes,
+            "__builtins__": {
+                "len": len,
+                "str": str,
+                "int": int,
+                "float": float,
+                "bool": bool,
+                "list": list,
+                "dict": dict,
+                "set": set,
+                "tuple": tuple,
+                "range": range,
+                "sum": sum,
+                "max": max,
+                "min": min,
+                "abs": abs,
+                "round": round,
+            },
+        }
+        
+        # Clean up query code - remove any import statements that might cause issues
+        # Split by lines and filter out import statements
+        lines = query_code.split("\n")
+        cleaned_lines = [
+            line
+            for line in lines
+            if not line.strip().startswith("import ") and not line.strip().startswith("from ")
+        ]
+        query_code = "\n".join(cleaned_lines)
+        
+        # Execute the query in a restricted namespace
+        local_vars = {}
+        exec(query_code, safe_globals, local_vars)  # noqa: S102
+        
+        # Get the result - check if 'result' variable exists, otherwise try 'df'
+        if "result" in local_vars:
+            result = local_vars["result"]
+        elif "df" in local_vars:
+            result = local_vars["df"]
+        else:
+            # If no explicit result variable, get the last expression
+            # This is a fallback - ideally the LLM should assign to 'result'
+            raise ValueError("Query must assign result to 'result' variable")
+        
+        # Check if a plot was generated
+        plot_image = None
+        if "plot_image" in local_vars:
+            plot_image = local_vars["plot_image"]
+            logger.info("Plot image generated")
+        
+        # Convert result to a serializable format
+        result = _convert_to_serializable(result)
+        
+        return {"result": result, "plot_image": plot_image}
+    except Exception as e:
+        logger.error("Error executing query: %s", e, exc_info=True)
+        # Provide more helpful error message
+        error_msg = str(e)
+        if "NaN" in error_msg or "NA" in error_msg:
+            error_msg = (
+                f"{error_msg}. "
+                "The query may need to handle missing values (NaN/NA) using .notna() or .dropna() before filtering."
+            )
+        raise ModelCannotRetry(f"Failed to execute query: {error_msg}") from e
+
+
+@last_model_retry_soft_fail
+async def data_analysis(ctx: RunContext, query: str) -> ToolReturn:
+    """
+    Analyze tabular data files (CSV, Excel) based on user query.
+    Can also generate plots/graphs/charts.
+    
+    This tool:
+    1. Loads the tabular file(s) from attachments
+    2. Generates metadata about the file structure
+    3. Uses an LLM to generate a pandas query based on user query
+    4. Executes the query and returns results
+    
+    Args:
+        ctx (RunContext): The run context containing the conversation.
+        query (str): The user's data analysis question.
+    
+    Returns:
+        ToolReturn: Contains the analysis results and metadata.
+    """
+    try:
+        # Find tabular files in attachments
+        # First, get all attachments for debugging
+        all_attachments = await sync_to_async(list)(
+            ctx.deps.conversation.attachments.all()
+        )
+        logger.info(
+            "All attachments in conversation: %s",
+            [
+                {
+                    "file_name": a.file_name,
+                    "content_type": a.content_type,
+                    "upload_state": a.upload_state,
+                    "conversion_from": a.conversion_from,
+                }
+                for a in all_attachments
+            ],
+        )
+
+        # Find tabular files - exclude converted files (they have conversion_from set)
+        # First try by content_type
+        tabular_attachments_by_type = await sync_to_async(list)(
+            ctx.deps.conversation.attachments.filter(
+                content_type__in=TABULAR_MIME_TYPES,
+                upload_state=AttachmentStatus.READY,
+            )
+            .filter(
+                Q(conversion_from__isnull=True) | Q(conversion_from="")
+            )
+        )
+        
+        # If no files found by content_type, try by file extension as fallback
+        # (some systems detect CSV as text/plain instead of text/csv)
+        if not tabular_attachments_by_type:
+            csv_extensions = [".csv", ".xlsx", ".xls"]
+            all_ready_attachments = await sync_to_async(list)(
+                ctx.deps.conversation.attachments.filter(
+                    upload_state=AttachmentStatus.READY,
+                )
+                .filter(
+                    Q(conversion_from__isnull=True) | Q(conversion_from="")
+                )
+            )
+            tabular_attachments = [
+                att
+                for att in all_ready_attachments
+                if any(att.file_name.lower().endswith(ext) for ext in csv_extensions)
+                # Exclude Markdown files (converted files have .md extension or content_type text/markdown)
+                and not att.file_name.lower().endswith(".md")
+                and att.content_type != "text/markdown"
+            ]
+            if tabular_attachments:
+                logger.info(
+                    "Found %d tabular file(s) by extension fallback (content_type was not recognized): %s",
+                    len(tabular_attachments),
+                    [f"{a.file_name} ({a.content_type})" for a in tabular_attachments],
+                )
+        else:
+            tabular_attachments = tabular_attachments_by_type
+
+        # If still no files found, check if there are converted files that might have originals
+        # This handles the case where an Excel file was converted to Markdown for RAG
+        if not tabular_attachments:
+            # Look for converted files with tabular extensions
+            csv_extensions = [".csv", ".xlsx", ".xls"]
+            converted_attachments = await sync_to_async(list)(
+                ctx.deps.conversation.attachments.filter(
+                    upload_state=AttachmentStatus.READY,
+                )
+                .exclude(
+                    Q(conversion_from__isnull=True) | Q(conversion_from="")
+                )
+            )
+            
+            # For each converted file, try to find the original
+            for converted_att in converted_attachments:
+                if any(converted_att.file_name.lower().endswith(ext) for ext in csv_extensions):
+                    # Try to find the original file using conversion_from key
+                    original_key = converted_att.conversion_from
+                    if original_key:
+                        original_attachment = await sync_to_async(
+                            ctx.deps.conversation.attachments.filter(
+                                key=original_key,
+                                upload_state=AttachmentStatus.READY,
+                            ).first
+                        )()
+                        if original_attachment:
+                            logger.info(
+                                "Found original file '%s' for converted file '%s'",
+                                original_attachment.file_name,
+                                converted_att.file_name,
+                            )
+                            tabular_attachments.append(original_attachment)
+                            break
+
+        logger.info(
+            "Found %d tabular attachment(s): %s",
+            len(tabular_attachments),
+            [f"{a.file_name} ({a.content_type})" for a in tabular_attachments],
+        )
+
+        if not tabular_attachments:
+            raise ModelCannotRetry(
+                "No tabular files (CSV or Excel) found in the conversation. "
+                "Please upload a CSV or Excel file first. "
+                "Note: If you uploaded an Excel file that was converted to Markdown for RAG, "
+                "the original file must still be available."
+            )
+
+        # Use the first tabular file
+        attachment = tabular_attachments[0]
+        logger.info("Analyzing file: %s (type: %s)", attachment.file_name, attachment.content_type)
+
+        # Load file data
+        file_data = await read_tabular_file(attachment)
+        
+        # Validate that this is actually a valid Excel/CSV file (not a converted Markdown file)
+        # Check if it's an Excel file that should have ZIP signature
+        if attachment.file_name.lower().endswith((".xlsx", ".xls", ".xlsm", ".xlsb")):
+            if not _is_valid_excel_file(file_data, attachment.file_name):
+                logger.warning(
+                    "File '%s' does not appear to be a valid Excel file. "
+                    "It may be a converted Markdown file. Searching for original...",
+                    attachment.file_name,
+                )
+                # Try to find the original file
+                # Look for an attachment with the same name but without conversion_from
+                original_attachment = await sync_to_async(
+                    ctx.deps.conversation.attachments.filter(
+                        file_name=attachment.file_name,
+                        upload_state=AttachmentStatus.READY,
+                    )
+                    .filter(
+                        Q(conversion_from__isnull=True) | Q(conversion_from="")
+                    )
+                    .exclude(pk=attachment.pk)
+                    .first
+                )()
+                
+                if original_attachment:
+                    logger.info(
+                        "Found original file '%s' (key: %s), using it instead",
+                        original_attachment.file_name,
+                        original_attachment.key,
+                    )
+                    attachment = original_attachment
+                    file_data = await read_tabular_file(attachment)
+                elif hasattr(attachment, 'conversion_from') and attachment.conversion_from:
+                    # Try to find by key if this file has a conversion_from
+                    original_attachment = await sync_to_async(
+                        ctx.deps.conversation.attachments.filter(
+                            key=attachment.conversion_from,
+                            upload_state=AttachmentStatus.READY,
+                        ).first
+                    )()
+                    if original_attachment:
+                        logger.info(
+                            "Found original file via conversion_from: '%s'",
+                            original_attachment.file_name,
+                        )
+                        attachment = original_attachment
+                        file_data = await read_tabular_file(attachment)
+                    else:
+                        raise ModelCannotRetry(
+                            f"File '{attachment.file_name}' appears to be a converted Markdown file, "
+                            "not the original Excel file. The original file is not available. "
+                            "Please re-upload the original Excel file."
+                        )
+                else:
+                    raise ModelCannotRetry(
+                        f"File '{attachment.file_name}' does not appear to be a valid Excel file. "
+                        "It may be corrupted or in an unsupported format."
+                    )
+
+        # Load into pandas DataFrames
+        dataframes = await load_dataframe(file_data, attachment.content_type, attachment.file_name)
+
+        # Generate metadata
+        metadata = generate_metadata(dataframes, attachment.file_name)
+        logger.debug("File metadata: %s", json.dumps(metadata, indent=2))
+
+        # Generate query using LLM
+        # NOTE:
+        # We intentionally create a "bare" Agent instance here instead of using BaseAgent
+        # with tools enabled. Using BaseAgent would attach all configured tools (including
+        # this data_analysis tool itself), which can cause the model to try to call tools
+        # while we're already inside a tool execution, leading to nested tool calls and
+        # failures like "Failed to generate query. Please try rephrasing your question.".
+        #
+        # Here we reuse the same model configuration as BaseAgent but WITHOUT any tools,
+        # so this internal call is purely text-to-text.
+        llm_config = settings.LLM_CONFIGURATIONS[settings.LLM_DEFAULT_MODEL_HRID]
+        if llm_config.is_custom:
+            model_instance = prepare_custom_model(llm_config)
+        else:
+            # Rely on pydantic-ai's built-in model registry / name inference
+            model_instance = llm_config.model_name
+
+        # Use the same keyword as when using BaseAgent, which forwards to Agent.
+        # On the current pydantic_ai version, the correct kwarg is `output_type`,
+        # not `result_type` (passing `result_type` raises a UserError).
+        query_agent = Agent(model=model_instance, output_type=str)
+        query_code = await generate_query(query, metadata, query_agent, ctx)
+        logger.debug("Generated query: %s", query_code)
+
+        # Execute query
+        try:
+            execution_result = await execute_query(query_code, dataframes)
+            result = execution_result.get("result")
+            plot_image_base64 = execution_result.get("plot_image")
+        except Exception as e:
+            logger.error("Query execution failed: %s", e, exc_info=True)
+            raise ModelRetry(
+                f"Failed to execute the generated query: {str(e)}. "
+                "Please try rephrasing your question."
+            ) from e
+
+        # Format result for return
+        return_value = {
+            "query": query,
+            "query_code": query_code,
+            "result": result,
+            "metadata": metadata,
+        }
+        
+        # Save plot image to storage if generated
+        plot_url = None
+        plot_attachment = None
+        if plot_image_base64:
+            try:
+                # Decode base64 image
+                plot_image_data = base64.b64decode(plot_image_base64)
+                
+                # Generate a unique filename for the plot
+                plot_filename = f"plot_{uuid.uuid4().hex[:8]}.png"
+                plot_key = f"{ctx.deps.conversation.pk}/plots/{plot_filename}"
+                
+                # Save to storage
+                await sync_to_async(default_storage.save)(
+                    plot_key, BytesIO(plot_image_data)
+                )
+                
+                # Create a permanent attachment record in the database
+                plot_attachment = await sync_to_async(ChatConversationAttachment.objects.create)(
+                    conversation=ctx.deps.conversation,
+                    uploaded_by=ctx.deps.user,
+                    key=plot_key,
+                    file_name=plot_filename,
+                    content_type="image/png",
+                    upload_state=AttachmentStatus.READY,
+                    size=len(plot_image_data),
+                )
+                
+                # Generate presigned URL for immediate access (valid for 1 hour)
+                plot_url = await sync_to_async(generate_retrieve_policy)(plot_key)
+                logger.info(
+                    "Plot image saved to storage and database: %s (presigned URL: %s)", 
+                    plot_key, 
+                    plot_url[:50] + "..."
+                )
+            except Exception as e:
+                logger.error("Failed to save plot image: %s", e, exc_info=True)
+                # Continue without plot URL if save fails
+        
+        if plot_url:
+            # Include both local and presigned URLs
+            return_value["plot_url"] = plot_url  # Presigned URL for direct access
+            return_value["plot_local_url"] = f"/media-key/{plot_key}"  # Local URL for reference
+            # Include attachment ID for reference
+            if plot_attachment:
+                return_value["plot_attachment_id"] = str(plot_attachment.pk)
+        
+        return ToolReturn(
+            return_value=return_value,
+            metadata={"file_name": attachment.file_name, "content_type": attachment.content_type},
+        )
+
+    except (ModelCannotRetry, ModelRetry):
+        # Re-raise these as-is
+        raise
+    except Exception as exc:
+        # Unexpected error - stop and inform user
+        logger.exception("Unexpected error in data_analysis: %s", exc)
+        raise ModelCannotRetry(
+            f"An unexpected error occurred during data analysis: {type(exc).__name__}. "
+            "You must explain this to the user and not try to answer based on your knowledge."
+        ) from exc
+
+
+def add_data_analysis_tool(agent: Agent) -> None:
+    """Add the data analysis tool to an existing agent."""
+
+    @agent.tool(retries=2)
+    @functools.wraps(data_analysis)
+    async def data_analysis_tool(ctx: RunContext, query: str) -> ToolReturn:
+        """
+        Analyze tabular data files (CSV, Excel) based on user query.
+        
+        This tool loads tabular files, generates metadata about their structure,
+        uses an LLM to generate a pandas query based on the user's question,
+        executes the query, and returns the results.
+        
+        Use this tool when the user asks questions about data in CSV or Excel files,
+        such as:
+        - "What is the average sales by region?"
+        - "Show me the top 10 products by revenue"
+        - "How many records are in this file?"
+        - "Filter data where column X is greater than Y"
+        
+        Args:
+            ctx (RunContext): The run context containing the conversation.
+            query (str): The user's data analysis question.
+        """
+        # Import here to avoid circular import
+        from chat.tools.data_analysis import data_analysis as _data_analysis
+        
+        return await _data_analysis(ctx, query)
+
+    @agent.instructions
+    def data_analysis_instructions() -> str:
+        """Dynamic system prompt function to add data analysis instructions."""
+        return (
+            "When the user asks questions about data in CSV or Excel files, "
+            "use the data_analysis tool to analyze the data and answer their question. "
+            "The tool will handle loading the file, generating queries, and executing them. "
+            "When a plot is generated, the tool returns a 'plot_url' in the result. "
+            "Use this presigned URL directly in markdown image syntax: ![Description](plot_url). "
+            "Do NOT use local URLs like /media-key/... - always use the presigned URL from plot_url. "
+            "Present the results clearly to the user."
+        )
+

src/backend/chat/tools/document_search_rag.py

@@ -39,7 +39,7 @@ def add_document_rag_search_tool(agent: Agent) -> None:
             metadata={"sources": {result.url for result in rag_results.data}},
         )
 
-    @agent.system_prompt
+    @agent.instructions
     def document_rag_instructions() -> str:
         """Dynamic system prompt function to add RAG instructions if any."""
         return (

src/backend/chat/tools/exceptions.py

@@ -3,17 +3,6 @@
 from pydantic_ai import ModelRetry
 
 
-class ModelRetryLast(ModelRetry):
-    """
-    Same as ModelRetry but also holds the last retry message to return when all attempts failed.
-    """
-
-    def __init__(self, message: str, last_retry_message: str):
-        """Initialize ModelRetryLast with message and last retry message."""
-        self.last_retry_message = last_retry_message
-        super().__init__(message)
-
-
 class ModelCannotRetry(ModelRetry):
     """
     Exception to raise when a tool function cannot be retried.

src/backend/chat/views.py

@@ -221,7 +221,7 @@ class ChatViewSet(  # pylint: disable=too-many-ancestors, abstract-method
         url_path="stop-streaming",
         url_name="stop-streaming",
     )
-    def post_stop_steaming(self, request, pk):  # pylint: disable=unused-argument
+    def post_stop_streaming(self, request, pk):  # pylint: disable=unused-argument
         """Handle POST requests to stop streaming the chat conversation.
 
         This action will put a poison pill in the redis cache to stop any ongoing streaming.

src/backend/core/api/__init__.py

@@ -1,12 +1,9 @@
 """Conversations core API endpoints"""
 
-from django.conf import settings
 from django.core.exceptions import ValidationError
 
 from rest_framework import exceptions as drf_exceptions
 from rest_framework import views as drf_views
-from rest_framework.decorators import api_view
-from rest_framework.response import Response
 
 
 def exception_handler(exc, context):
@@ -28,14 +25,3 @@ def exception_handler(exc, context):
         exc = drf_exceptions.ValidationError(detail=detail)
 
     return drf_views.exception_handler(exc, context)
-
-
-# pylint: disable=unused-argument
-@api_view(["GET"])
-def get_frontend_configuration(request):
-    """Returns the frontend configuration dict as configured in settings."""
-    frontend_configuration = {
-        "LANGUAGE_CODE": settings.LANGUAGE_CODE,
-    }
-    frontend_configuration.update(settings.FRONTEND_CONFIGURATION)
-    return Response(frontend_configuration)

src/backend/core/api/serializers.py

@@ -20,23 +20,3 @@ class UserSerializer(serializers.ModelSerializer):
             "sub",
         ]
         read_only_fields = ["id", "email", "full_name", "short_name", "sub"]
-
-
-class UserLightSerializer(UserSerializer):
-    """Serialize users with limited fields."""
-
-    id = serializers.SerializerMethodField(read_only=True)
-    email = serializers.SerializerMethodField(read_only=True)
-
-    def get_id(self, _user):
-        """Return always None. Here to have the same fields than in UserSerializer."""
-        return None
-
-    def get_email(self, _user):
-        """Return always None. Here to have the same fields than in UserSerializer."""
-        return None
-
-    class Meta:
-        model = models.User
-        fields = ["id", "email", "full_name", "short_name"]
-        read_only_fields = ["id", "email", "full_name", "short_name"]

src/backend/core/authentication/__init__.py

@@ -1,52 +0,0 @@
-"""Custom authentication classes for the Conversations core app"""
-
-from django.conf import settings
-
-from rest_framework.authentication import BaseAuthentication
-from rest_framework.exceptions import AuthenticationFailed
-
-
-class ServerToServerAuthentication(BaseAuthentication):
-    """
-    Custom authentication class for server-to-server requests.
-    Validates the presence and correctness of the Authorization header.
-    """
-
-    AUTH_HEADER = "Authorization"
-    TOKEN_TYPE = "Bearer"  # noqa S105
-
-    def authenticate(self, request):
-        """
-        Authenticate the server-to-server request by validating the Authorization header.
-
-        This method checks if the Authorization header is present in the request, ensures it
-        contains a valid token with the correct format, and verifies the token against the
-        list of allowed server-to-server tokens. If the header is missing, improperly formatted,
-        or contains an invalid token, an AuthenticationFailed exception is raised.
-
-        Returns:
-            None: If authentication is successful
-                  (no user is authenticated for server-to-server requests).
-
-        Raises:
-            AuthenticationFailed: If the Authorization header is missing, malformed,
-            or contains an invalid token.
-        """
-        auth_header = request.headers.get(self.AUTH_HEADER)
-        if not auth_header:
-            raise AuthenticationFailed("Authorization header is missing.")
-
-        # Validate token format and existence
-        auth_parts = auth_header.split(" ")
-        if len(auth_parts) != 2 or auth_parts[0] != self.TOKEN_TYPE:
-            raise AuthenticationFailed("Invalid authorization header.")
-
-        token = auth_parts[1]
-        if token not in settings.SERVER_TO_SERVER_API_TOKENS:
-            raise AuthenticationFailed("Invalid server-to-server token.")
-
-        # Authentication is successful, but no user is authenticated
-
-    def authenticate_header(self, request):
-        """Return the WWW-Authenticate header value."""
-        return f"{self.TOKEN_TYPE} realm='Create document server to server'"

src/backend/core/fields.py

@@ -1,25 +0,0 @@
-"""A JSONField for DRF to handle serialization/deserialization."""
-
-import json
-
-from rest_framework import serializers
-
-
-class JSONField(serializers.Field):
-    """
-    A custom field for handling JSON data.
-    """
-
-    def to_representation(self, value):
-        """
-        Convert the JSON string to a Python dictionary for serialization.
-        """
-        return value
-
-    def to_internal_value(self, data):
-        """
-        Convert the Python dictionary to a JSON string for deserialization.
-        """
-        if data is None:
-            return None
-        return json.dumps(data)

src/backend/core/filters.py

@@ -2,31 +2,9 @@
 
 import unicodedata
 
-import django_filters
-
 
 def remove_accents(value):
     """Remove accents from a string (vélo -> velo)."""
     return "".join(
         c for c in unicodedata.normalize("NFD", value) if unicodedata.category(c) != "Mn"
     )
-
-
-class AccentInsensitiveCharFilter(django_filters.CharFilter):
-    """
-    A custom CharFilter that filters on the accent-insensitive value searched.
-    """
-
-    def filter(self, qs, value):
-        """
-        Apply the filter to the queryset using the unaccented version of the field.
-
-        Args:
-            qs: The queryset to filter.
-            value: The value to search for in the unaccented field.
-        Returns:
-            A filtered queryset.
-        """
-        if value:
-            value = remove_accents(value)
-        return super().filter(qs, value)

src/backend/core/templates/core/generate_document.html

@@ -1,14 +0,0 @@
-<!DOCTYPE html>
-<html>
-<head>
-    <title>Generate Document</title>
-</head>
-<body>
-    <h2>Generate Document</h2>
-    <form method="post" enctype="multipart/form-data">
-        {% csrf_token %}
-        {{ form.as_p }}
-        <button type="submit">Generate PDF</button>
-    </form>
-</body>
-</html>

src/backend/core/templatetags/extra_tags.py

@@ -1,58 +0,0 @@
-"""Custom template tags for the core application of People."""
-
-import base64
-
-from django import template
-from django.contrib.staticfiles import finders
-
-from PIL import ImageFile as PillowImageFile
-
-register = template.Library()
-
-
-def image_to_base64(file_or_path, close=False):
-    """
-    Return the src string of the base64 encoding of an image represented by its path
-    or file opened or not.
-
-    Inspired by Django's "get_image_dimensions"
-    """
-    pil_parser = PillowImageFile.Parser()
-    if hasattr(file_or_path, "read"):
-        file = file_or_path
-        if file.closed and hasattr(file, "open"):
-            file_or_path.open()
-        file_pos = file.tell()
-        file.seek(0)
-    else:
-        try:
-            # pylint: disable=consider-using-with
-            file = open(file_or_path, "rb")
-        except OSError:
-            return ""
-        close = True
-
-    try:
-        image_data = file.read()
-        if not image_data:
-            return ""
-        pil_parser.feed(image_data)
-        if pil_parser.image:
-            mime_type = pil_parser.image.get_format_mimetype()
-            encoded_string = base64.b64encode(image_data)
-            return f"data:{mime_type:s};base64, {encoded_string.decode('utf-8'):s}"
-        return ""
-    finally:
-        if close:
-            file.close()
-        else:
-            file.seek(file_pos)
-
-
-@register.simple_tag
-def base64_static(path):
-    """Return a static file into a base64."""
-    full_path = finders.find(path)
-    if full_path:
-        return image_to_base64(full_path, True)
-    return ""

src/backend/pyproject.toml

@@ -53,6 +53,10 @@ dependencies = [
     "markitdown==0.0.2",
     "mozilla-django-oidc==4.0.1",
     "nested-multipart-parser==1.6.0",
+    "matplotlib==3.9.2",
+    "numpy==2.1.3",
+    "openpyxl==3.1.5",
+    "pandas==2.2.3",
     "posthog==7.0.0",
     "pydantic==2.12.4",
     "pydantic-ai-slim[openai,mistral,mcp,evals,logfire]==1.17.0",

src/frontend/apps/conversations/src/components/Loader.tsx

@@ -1,8 +1,7 @@
 import dynamic from 'next/dynamic';
 
-import searchingAnimation from '@/assets/lotties/searching';
-
 const Lottie = dynamic(() => import('lottie-react'), { ssr: false });
+import searchingAnimation from '@/assets/lotties/searching';
 
 export function Loader() {
   return (

src/frontend/apps/conversations/src/features/chat/components/ChatMessage.tsx

@@ -1,386 +0,0 @@
-import {
-  Message,
-  ReasoningUIPart,
-  SourceUIPart,
-  ToolInvocationUIPart,
-} from '@ai-sdk/ui-utils';
-import 'katex/dist/katex.min.css';
-import { memo, useDeferredValue } from 'react';
-import { useTranslation } from 'react-i18next';
-import { MarkdownHooks } from 'react-markdown';
-import rehypeKatex from 'rehype-katex';
-import rehypePrettyCode from 'rehype-pretty-code';
-import remarkGfm from 'remark-gfm';
-import remarkMath from 'remark-math';
-
-import { Box, Icon, Text } from '@/components';
-import { useClipboard } from '@/hook';
-import { useResponsiveStore } from '@/stores';
-
-import { AttachmentList } from './AttachmentList';
-import { CodeBlock } from './CodeBlock';
-import { FeedbackButtons } from './FeedbackButtons';
-import { SourceItemList } from './SourceItemList';
-import { ToolInvocationItem } from './ToolInvocationItem';
-
-// Mémoriser les plugins Markdown en dehors du composant pour éviter les recréations
-const remarkPlugins = [remarkGfm, remarkMath];
-const rehypePlugins = [
-  [
-    rehypePrettyCode,
-    {
-      theme: 'github-dark-dimmed',
-    },
-  ],
-  rehypeKatex,
-];
-
-// Composants Markdown mémorisés
-// eslint-disable-next-line @typescript-eslint/no-explicit-any
-const markdownComponents: any = {
-  // eslint-disable-next-line @typescript-eslint/no-unused-vars, @typescript-eslint/no-explicit-any
-  p: ({ node, ...props }: any) => (
-    <Text
-      as="p"
-      $css="display: block"
-      $theme="greyscale"
-      $variation="850"
-      {...props}
-    />
-  ),
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  a: ({ children, ...props }: any) => (
-    <a target="_blank" rel="noopener noreferrer" {...props}>
-      {children}
-    </a>
-  ),
-  // eslint-disable-next-line @typescript-eslint/no-unused-vars, @typescript-eslint/no-explicit-any
-  pre: ({ node, children, ...props }: any) => (
-    <CodeBlock {...props}>{children}</CodeBlock>
-  ),
-};
-
-// Composant Markdown mémorisé pour éviter les recalculs inutiles
-const MemoizedMarkdown = memo(function MemoizedMarkdown({
-  content,
-}: {
-  content: string;
-}) {
-  return (
-    <MarkdownHooks
-      remarkPlugins={remarkPlugins}
-      // eslint-disable-next-line @typescript-eslint/no-explicit-any, @typescript-eslint/no-unsafe-assignment
-      rehypePlugins={rehypePlugins as any} // Type mismatch with react-markdown types
-      // eslint-disable-next-line @typescript-eslint/no-unsafe-assignment
-      components={markdownComponents}
-    >
-      {content}
-    </MarkdownHooks>
-  );
-});
-
-interface ChatMessageProps {
-  message: Message;
-  isLastAssistantMessageInConversation: boolean;
-  shouldApplyStreamingHeight: boolean;
-  streamingMessageHeight: number | null;
-  isCurrentlyStreaming: boolean;
-  status: 'idle' | 'streaming' | 'submitted' | 'ready' | 'error';
-  isSourceOpen: string | null;
-  conversationId: string | undefined;
-  onOpenSources: (messageId: string) => void;
-  getMetadata: (url: string) =>
-    | {
-        title: string | null;
-        favicon: string | null;
-        loading: boolean;
-        error: boolean;
-      }
-    | undefined;
-}
-
-export const ChatMessage = memo(function ChatMessage({
-  message,
-  isLastAssistantMessageInConversation,
-  shouldApplyStreamingHeight,
-  streamingMessageHeight,
-  isCurrentlyStreaming,
-  status,
-  isSourceOpen,
-  conversationId,
-  onOpenSources,
-  getMetadata,
-}: ChatMessageProps) {
-  const { t } = useTranslation();
-  const copyToClipboard = useClipboard();
-  const { isMobile } = useResponsiveStore();
-
-  const deferredContent = useDeferredValue(message.content);
-
-  const contentToRender =
-    message.role === 'assistant' ? deferredContent : message.content;
-
-  return (
-    <Box
-      key={message.id}
-      data-message-id={message.id}
-      $css={`
-        display: flex;
-        width: 100%;
-        margin: auto;
-        margin-bottom: ${isLastAssistantMessageInConversation ? '30px' : '0px'};
-        color: var(--c--theme--colors--greyscale-850);
-        padding-left: 12px;
-        padding-right: 12px;
-        max-width: 750px;
-        text-align: left;
-        overflow-wrap: anywhere;
-        flex-direction: ${message.role === 'user' ? 'row-reverse' : 'row'};
-      `}
-    >
-      <Box
-        $display="block"
-        $width={`${message.role === 'user' ? 'auto' : '100%'}`}
-      >
-        {message.experimental_attachments &&
-          message.experimental_attachments.length > 0 && (
-            <Box>
-              <AttachmentList
-                attachments={message.experimental_attachments}
-                isReadOnly={true}
-              />
-            </Box>
-          )}
-        <Box
-          $radius="8px"
-          $width={`${message.role === 'user' ? 'auto' : '100%'}`}
-          $maxWidth="100%"
-          $padding={`${message.role === 'user' ? '12px' : '0'}`}
-          $margin={{ vertical: 'base' }}
-          $background={`${message.role === 'user' ? '#EEF1F4' : 'white'}`}
-          $css={`
-            display: inline-block;
-            float: right;
-            ${shouldApplyStreamingHeight ? `min-height: ${streamingMessageHeight}px;` : ''}
-          `}
-        >
-          {message.content && (
-            <Box
-              className="mainContent-chat"
-              data-testid={
-                message.role === 'assistant'
-                  ? 'assistant-message-content'
-                  : undefined
-              }
-              $padding={{ all: 'xxs' }}
-            >
-              <p className="sr-only">
-                {message.role === 'user'
-                  ? t('You said: ')
-                  : t('Assistant IA replied: ')}
-              </p>
-              {message.role === 'user' ? (
-                <Text
-                  as="p"
-                  $css="white-space: pre-wrap; display: block;"
-                  $theme="greyscale"
-                  $variation="850"
-                >
-                  {message.content}
-                </Text>
-              ) : (
-                <MemoizedMarkdown content={contentToRender} />
-              )}
-            </Box>
-          )}
-
-          <Box $direction="column" $gap="2">
-            {isCurrentlyStreaming &&
-              isLastAssistantMessageInConversation &&
-              status === 'streaming' &&
-              message.parts?.some(
-                (part) =>
-                  part.type === 'tool-invocation' &&
-                  part.toolInvocation.toolName !== 'document_parsing',
-              ) && (
-                <Box
-                  $direction="row"
-                  $align="center"
-                  $gap="6px"
-                  $width="100%"
-                  $maxWidth="750px"
-                  $margin={{
-                    all: 'auto',
-                    top: 'base',
-                    bottom: 'md',
-                  }}
-                >
-                  <Text $variation="600" $size="md">
-                    {(() => {
-                      const toolInvocation = message.parts?.find(
-                        (part) =>
-                          part.type === 'tool-invocation' &&
-                          part.toolInvocation.toolName !== 'document_parsing',
-                      );
-                      if (
-                        toolInvocation?.type === 'tool-invocation' &&
-                        toolInvocation.toolInvocation.toolName === 'summarize'
-                      ) {
-                        return t('Summarizing...');
-                      }
-                      return t('Search...');
-                    })()}
-                  </Text>
-                </Box>
-              )}
-            {message.parts
-              ?.filter(
-                (part) =>
-                  part.type === 'reasoning' || part.type === 'tool-invocation',
-              )
-              .map(
-                (
-                  part: ReasoningUIPart | ToolInvocationUIPart,
-                  partIndex: number,
-                ) =>
-                  part.type === 'reasoning' ? (
-                    <Box
-                      key={`reasoning-${partIndex}`}
-                      $background="var(--c--theme--colors--greyscale-100)"
-                      $color="var(--c--theme--colors--greyscale-500)"
-                      $padding={{ all: 'sm' }}
-                      $radius="md"
-                      $css="font-size: 0.9em;"
-                    >
-                      {part.reasoning}
-                    </Box>
-                  ) : part.type === 'tool-invocation' &&
-                    isCurrentlyStreaming &&
-                    isLastAssistantMessageInConversation ? (
-                    <ToolInvocationItem
-                      key={`tool-invocation-${partIndex}`}
-                      toolInvocation={part.toolInvocation}
-                      status={status}
-                      hideSearchLoader={true}
-                    />
-                  ) : null,
-              )}
-          </Box>
-          {message.role === 'assistant' &&
-            !(
-              isLastAssistantMessageInConversation && status === 'streaming'
-            ) && (
-              <Box
-                $css="color: #222631; font-size: 12px;"
-                $direction="row"
-                $align="center"
-                $justify="space-between"
-                $gap="6px"
-                $margin={{ top: 'base' }}
-              >
-                <Box $direction="row" $gap="4px">
-                  <Box
-                    $direction="row"
-                    $align="center"
-                    $gap="4px"
-                    className="c__button--neutral action-chat-button"
-                    onClick={() => copyToClipboard(message.content)}
-                    onKeyDown={(e) => {
-                      if (e.key === 'Enter' || e.key === ' ') {
-                        e.preventDefault();
-                        copyToClipboard(message.content);
-                      }
-                    }}
-                    role="button"
-                    tabIndex={0}
-                  >
-                    <Icon
-                      iconName="content_copy"
-                      $theme="greyscale"
-                      $variation="550"
-                      $size="16px"
-                      className="action-chat-button-icon"
-                    />
-                    {!isMobile && (
-                      <Text $theme="greyscale" $variation="550">
-                        {t('Copy')}
-                      </Text>
-                    )}
-                  </Box>
-                  {message.parts?.some((part) => part.type === 'source') &&
-                    (() => {
-                      const sourceCount =
-                        message.parts?.filter((part) => part.type === 'source')
-                          .length || 0;
-                      return (
-                        <Box
-                          $direction="row"
-                          $align="center"
-                          $gap="4px"
-                          className={`c__button--neutral action-chat-button ${isSourceOpen === message.id ? 'action-chat-button--open' : ''}`}
-                          onClick={() => onOpenSources(message.id)}
-                          onKeyDown={(e) => {
-                            if (e.key === 'Enter' || e.key === ' ') {
-                              e.preventDefault();
-                              onOpenSources(message.id);
-                            }
-                          }}
-                          role="button"
-                          tabIndex={0}
-                        >
-                          <Icon
-                            iconName="book"
-                            $theme="greyscale"
-                            $variation="550"
-                            $size="16px"
-                            className="action-chat-button-icon"
-                          />
-                          <Text
-                            $theme="greyscale"
-                            $variation="550"
-                            $weight="500"
-                            $size="12px"
-                          >
-                            {t('Show')} {sourceCount}{' '}
-                            {sourceCount !== 1 ? t('sources') : t('source')}
-                          </Text>
-                        </Box>
-                      );
-                    })()}
-                </Box>
-                <Box $direction="row" $gap="4px">
-                  {conversationId &&
-                    message.id &&
-                    message.id.startsWith('trace-') && (
-                      <FeedbackButtons
-                        conversationId={conversationId}
-                        messageId={message.id}
-                      />
-                    )}
-                </Box>
-              </Box>
-            )}
-          {message.parts &&
-            isSourceOpen === message.id &&
-            (() => {
-              const sourceParts = message.parts.filter(
-                (part): part is SourceUIPart => part.type === 'source',
-              );
-              return (
-                <Box
-                  $css={`
-                    animation: fade-in 0.2s ease-out;
-                  `}
-                >
-                  <SourceItemList
-                    parts={sourceParts}
-                    getMetadata={getMetadata}
-                  />
-                </Box>
-              );
-            })()}
-        </Box>
-      </Box>
-    </Box>
-  );
-});

src/frontend/apps/conversations/src/features/chat/hooks/index.ts

@@ -1,3 +1,2 @@
 export { useChatScroll } from './useChatScroll';
 export { useSourceMetadataCache } from './useSourceMetadata';
-export { useModelSelection } from './useModelSelection';

src/frontend/apps/conversations/src/features/chat/hooks/useModelSelection.ts

@@ -1,44 +0,0 @@
-import { useEffect, useRef, useState } from 'react';
-
-import { LLMModel, useLLMConfiguration } from '../api/useLLMConfiguration';
-import { useChatPreferencesStore } from '../stores/useChatPreferencesStore';
-
-export const useModelSelection = () => {
-  const { data: llmConfig } = useLLMConfiguration();
-  const { selectedModelHrid, setSelectedModelHrid } = useChatPreferencesStore();
-  const [selectedModel, setSelectedModel] = useState<LLMModel | null>(null);
-  const hasInitializedRef = useRef(false);
-
-  useEffect(() => {
-    // Ne s'exécuter qu'une seule fois quand llmConfig est chargé
-    if (llmConfig?.models && !hasInitializedRef.current) {
-      let modelToSelect: LLMModel | undefined;
-
-      if (selectedModelHrid) {
-        // Try to find the previously selected model
-        modelToSelect = llmConfig.models.find(
-          (model) =>
-            model.hrid === selectedModelHrid && model.is_active !== false,
-        );
-      }
-
-      // If no saved model or saved model not found/inactive, use default
-      if (!modelToSelect) {
-        modelToSelect = llmConfig.models.find((model) => model.is_default);
-      }
-
-      if (modelToSelect) {
-        setSelectedModel(modelToSelect);
-        setSelectedModelHrid(modelToSelect.hrid);
-        hasInitializedRef.current = true;
-      }
-    }
-  }, [llmConfig?.models, selectedModelHrid, setSelectedModelHrid]);
-
-  const handleModelSelect = (model: LLMModel) => {
-    setSelectedModel(model);
-    setSelectedModelHrid(model.hrid);
-  };
-
-  return { selectedModel, handleModelSelect };
-};