✨(summarize) new summarize tool integration

Improve the existing tool to manage bigger documents.

CHANGELOG.md

@@ -13,6 +13,7 @@ and this project adheres to
 - 🐛(front) fix target blank links in chat #103
 - 🚑️(posthog) pass str instead of UUID for user PK #134
 - ⚡️(web-search) keep running when tool call fails #137
+- ✨(summarize): new summarize tool integration #78
 
 ### Removed
 

src/backend/chat/agents/summarize.py

@@ -2,6 +2,7 @@
 
 import dataclasses
 import logging
+import asyncio
 
 from django.conf import settings
 from django.core.files.storage import default_storage
@@ -11,6 +12,7 @@ from pydantic_ai import RunContext
 from pydantic_ai.messages import ToolReturn
 
 from .base import BaseAgent
+from ..tools.document_search_rag import add_document_rag_search_tool
 
 logger = logging.getLogger(__name__)
 
@@ -35,14 +37,18 @@ def read_document_content(doc):
         return doc.file_name, f.read().decode("utf-8")
 
 
-async def hand_off_to_summarization_agent(ctx: RunContext) -> ToolReturn:
+async def hand_off_to_summarization_agent(
+    ctx: RunContext, *, instructions: str | None = None
+) -> ToolReturn:
     """
-    Generate a complete, ready-to-use summary of the documents in context
-    (do not request the documents to the user).
-    Return this summary directly to the user WITHOUT any modification,
-    or additional summarization.
-    The summary is already optimized and MUST be presented as-is in the final response
-    or translated preserving the information.
+    Summarize the documents for the user, only when asked for.
+    Instructions are optional but should reflect the user's request.
+    Examples :
+    "Résume ce doc en 2 paragraphes" -> instructions = "résumé en 2 paragraphes"
+    "Résume ce doc en anglais" -> instructions = "In English"
+    "Résume ce doc" -> instructions = "" (default)
+    Args:
+        instructions (str | None): The instructions the user gave to use for the summarization
     """
     summarization_agent = SummarizationAgent()
 
@@ -53,6 +59,8 @@ async def hand_off_to_summarization_agent(ctx: RunContext) -> ToolReturn:
         "Document contents:\n"
         "{documents_prompt}\n"
     )
+
+    # Collect documents content
     text_attachment = await sync_to_async(list)(
         ctx.deps.conversation.attachments.filter(
             content_type__startswith="text/",
@@ -61,25 +69,72 @@ async def hand_off_to_summarization_agent(ctx: RunContext) -> ToolReturn:
 
     documents = [await read_document_content(doc) for doc in text_attachment]
 
-    documents_prompt = "\n\n".join(
-        [
-            (f"<document>\n<name>\n{name}\n</name>\n<content>\n{content}\n</content>\n</document>")
-            for name, content in documents
+    # Instructions: rely on tool argument only; model should extract them upstream
+    if instructions is not None:
+        instructions_hint: str = instructions.strip()
+    else:
+        instructions_hint = ""
+
+    # Helpers
+    def chunk_text(text: str, size: int = 10000) -> list[str]:
+        if size <= 0:
+            return [text]
+        return [text[i : i + size] for i in range(0, len(text), size)]
+
+    # 2) Chunk documents and summarize each chunk
+    full_text = "\n\n".join(doc[1] for doc in documents)
+    chunks = chunk_text(full_text, size=10000)
+    logger.info(
+        "[summarize] chunking: %s parts (size~%s), instructions='%s'",
+        len(chunks),
+        10000,
+        instructions_hint or "",
+    )
+
+    async def summarize_chunk(idx, chunk, total_chunks, summarization_agent, ctx):
+        sum_prompt = (
+            "Tu es un agent spécialisé en synthèses de textes. "
+            "Génère un résumé clair et concis du passage suivant (partie {idx}/{total}) :\n"
+            "'''\n{context}\n'''\n\n"
+        ).format(context=chunk, idx=idx, total=total_chunks)
+        logger.info("[summarize] CHUNK %s/%s prompt=> %s", idx, total_chunks, sum_prompt[0:100]+'...')
+        resp = await summarization_agent.run(sum_prompt, usage=ctx.usage)
+        logger.info("[summarize] CHUNK %s/%s response<= %s", idx, total_chunks, resp.output or "")
+        return resp.output or ""
+
+    # Parallelize the chunk summarization in batches of 5 using asyncio.gather
+    chunk_summaries: list[str] = []
+    batch_size = 5
+    for start_idx in range(0, len(chunks), batch_size):
+        end_idx = start_idx + batch_size
+        batch_chunks = chunks[start_idx:end_idx]
+        summarization_tasks = [
+            summarize_chunk(idx, chunk, len(chunks), summarization_agent, ctx)
+            for idx, chunk in enumerate(batch_chunks, start=start_idx + 1)
         ]
-    )
+        batch_results = await asyncio.gather(*summarization_tasks)
+        chunk_summaries.extend(batch_results)
 
-    formatted_prompt = prompt.format(
-        user_prompt=ctx.prompt,
-        documents_prompt=documents_prompt,
-    )
+    if not instructions_hint:
+        instructions_hint = "Le résumé doit être en Français, contenir 2 ou 3 parties."
 
-    logger.debug("Summarize prompt: %s", formatted_prompt)
-
-    response = await summarization_agent.run(formatted_prompt, usage=ctx.usage)
-
-    logger.debug("Summarize response: %s", response)
+    # 3) Merge chunk summaries into a single concise summary
+    merged_prompt = (
+        "Produit une synthèse cohérente à partir des résumés ci-dessous.\n\n"
+        "'''\n{context}\n'''\n\n"
+        "Contraintes :\n"
+        "- Résumer sans répéter.\n"
+        "- Harmoniser le style et la terminologie.\n"
+        "- Le résumé final doit être bien structuré et formaté en markdown. \n"
+        "- Respecter les consignes : {instructions}\n"
+        "Réponds directement avec le résumé final."
+    ).format(context="\n\n".join(chunk_summaries), instructions=instructions_hint or "")
+    logger.info("[summarize] MERGE prompt=> %s", merged_prompt)
+    merged_resp = await summarization_agent.run(merged_prompt, usage=ctx.usage)
+    final_summary = (merged_resp.output or "").strip()
+    logger.info("[summarize] MERGE response<= %s", final_summary)
 
     return ToolReturn(
-        return_value=response.output,
+        return_value=final_summary,
         metadata={"sources": {doc[0] for doc in documents}},
     )

src/backend/chat/clients/pydantic_ai.py

@@ -483,14 +483,20 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
             @self.conversation_agent.system_prompt
             def summarization_system_prompt() -> str:
                 return (
-                    "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."
+                    "When the user asks to summarize attached document(s), you MUST call the"
+                    " summarize tool. Pass user's instructions if provided, otherwise pass an"
+                    " empty instructions string once the user confirms (e.g. says 'ok'). Do NOT"
+                    " call web search or document_search_rag to produce summaries; rely only on"
+                    " the attached documents stored in context."
+                )
+
+            # Inform the model (system-level) that documents are attached and available
+            @self.conversation_agent.system_prompt
+            def attached_documents_note() -> str:
+                return (
+                    "[Internal context] User documents are attached to this conversation. "
+                    "Do not request re-upload of documents; consider them already available "
+                    "via the internal store."
                 )
 
             @self.conversation_agent.tool

src/backend/chat/tools/document_search_rag.py

@@ -20,8 +20,12 @@ def add_document_rag_search_tool(agent: Agent) -> None:
 
         Args:
             ctx (RunContext): The run context containing the conversation.
-            query (str): The term to search the internet for.
+            query (str): The query to search the documents for.
         """
+        # Defensive: ctx.deps or ctx.deps.conversation may be unavailable in some flows (start of conversation)
+        if not getattr(ctx, "deps", None) or not getattr(ctx.deps, "conversation", None):
+            return ToolReturn(return_value=[], content="", metadata={"sources": set()})
+
         document_store_backend = import_string(settings.RAG_DOCUMENT_SEARCH_BACKEND)
 
         document_store = document_store_backend(ctx.deps.conversation.collection_id)
@@ -43,8 +47,6 @@ def add_document_rag_search_tool(agent: Agent) -> None:
     def document_rag_instructions() -> str:
         """Dynamic system prompt function to add RAG instructions if any."""
         return (
-            "If the user wants specific information from a document, invoke "
-            "web_search_albert_rag with an appropriate query string."
-            "Do not ask the user for the document; rely on the tool to locate "
-            "and return relevant passages."
+            "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."
         )

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

@@ -68,7 +68,7 @@ export const ToolInvocationItem: React.FC<ToolInvocationItemProps> = ({
         >
           <Loader />
           <Text $variation="600" $size="md">
-            {t('Search...')}
+            {toolInvocation.toolName === 'summarize' ? t('Summarizing...') : t('Search...')}
           </Text>
         </Box>
       )}