- MISTRAL_SENTINELLE_GUIDE.md: health monitoring, weekly benchmarks - MISTRAL_TOWER_GUIDE.md: knowledge RAG, commercial docs - MISTRAL_FORGE_GUIDE.md: Codestral FIM, dataset pipeline, fine-tune - MISTRAL_DEVSTRAL_GUIDE.md: 4 engineering profiles, dispatch - cron_model_audit.sh: weekly 10-prompt audit, baseline comparison, alerts Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
5.5 KiB
Mistral Tower Guide
How Tower manages knowledge in the Kill_LIFE / Mascarade ecosystem
Agent: Tower
Agent ID: ag_019d124e760877359ad3ff5031179ebc
Model: magistral-medium-latest (temperature 0.4)
Domains: docs, readme, specs, content, email, crm, training, research
Overview
Tower is the knowledge and content agent of the Mascarade mesh. Its primary responsibilities are:
- Knowledge management via the knowledge-base MCP server (Memos + Docmost backends)
- RAG-assisted document retrieval using Mistral Document Library (Beta Libraries)
- Commercial content generation (formation docs, slide decks, product sheets)
- Research and analysis tasks (veille technologique, market analysis)
Tower operates with higher creativity (temperature 0.4) than other agents, tuned for long-form text generation and nuanced content production.
Knowledge Base MCP Server
Spec: specs/knowledge_base_mcp_spec.md
MCP config: mcp.json (entry knowledge-base)
Runner: tools/run_knowledge_base_mcp.sh
Backend: mascarade/core/mascarade/integrations/knowledge_base.py
Architecture
The knowledge-base MCP server bridges two wiki/note backends:
| Backend | Role | URL |
|---|---|---|
| Memos | Quick notes, operational memos | Self-hosted |
| Docmost | Structured documentation, long-form wiki | Self-hosted |
MCP tools exposed
| Tool | Description |
|---|---|
search_pages |
Full-text search across both backends |
read_page |
Retrieve a specific page by ID |
These tools are validated via test/test_knowledge_base_mcp.py which confirms server initialization and tool registration.
Integration with Mascarade core
The knowledge-base MCP is used by:
- Core routes:
knowledge-base/* - Knowledge scribe agent:
agents/knowledge-scribe/run-and-push - Tower agent: for RAG context injection before generating content
Mistral Document Library (RAG)
Task: T-MS-013 (Plan 24)
Client: mascarade/agents/mistral_agents_beta_api.py (MistralLibraryClient)
What it is
Mistral Document Library (Beta Libraries) is Mistral's native RAG solution. Documents uploaded via the Files API are associated with an agent and become searchable context.
How Tower uses it
- Commercial documents from
docs/commercial/are uploaded to Mistral Files API (T-MS-004) - Documents are associated with the Tower agent via Library configuration
- When Tower receives a query, it automatically searches the library for relevant context
- RAG-augmented responses include citations from uploaded documents
Commercial documents available
| Document | Path |
|---|---|
| Factory 4.0 Starter | docs/commercial/factory_4_0_starter.md |
| Factory 4.0 Pro | docs/commercial/factory_4_0_pro.md |
| Factory 4.0 Enterprise | docs/commercial/factory_4_0_enterprise.md |
| Factory 4.0 Slide Deck | docs/commercial/factory_4_0_slide_deck.md |
Status
- Documents exported to
docs/commercial/(done) - Upload to Mistral Files API: pending (T-MS-004, requires API call)
- Library association with Tower agent: pending (T-MS-013, depends on T-MS-004)
Content Generation Profiles
Tower operates in three content profiles, selectable via dispatch_to_agent.sh:
Writer profile (docs, readme, specs, content, wiki, markdown)
bash tools/ai/dispatch_to_agent.sh --lot T-MS-032 --domain docs
Tasks: documentation pages, README files, technical specifications, wiki content.
Commercial profile (email, crm, commercial, training, formation)
bash tools/ai/dispatch_to_agent.sh --lot T-MS-004 --domain commercial
Tasks: formation documents, product sheets, email templates, CRM content, lead scoring with RAG context.
Researcher profile (research, veille, analysis)
bash tools/ai/dispatch_to_agent.sh --lot T-MS-032 --domain research
Tasks: technology watch, competitor analysis, market research synthesis.
Dispatch via dispatch_to_agent.sh
Location: tools/ai/dispatch_to_agent.sh
# Documentation task
bash tools/ai/dispatch_to_agent.sh --lot T-MS-032 --domain docs --prompt "Draft the Mistral Studio Overview page"
# Commercial content with local model (zero cost)
bash tools/ai/dispatch_to_agent.sh --lot T-MS-004 --domain commercial --local
# Research task
bash tools/ai/dispatch_to_agent.sh --lot T-MS-032 --domain research
# Dry run to inspect prompt
bash tools/ai/dispatch_to_agent.sh --lot T-MS-032 --domain docs --dry-run
RAG Pipeline (future state)
Once T-MS-004 and T-MS-013 are completed, the Tower RAG pipeline will operate as follows:
User query
|
v
Tower agent (magistral-medium-latest, temp 0.4)
|
+-> Mistral Document Library search (RAG)
| |
| +-> docs/commercial/* (uploaded PDFs/markdown)
| +-> Outline wiki pages (if connected)
|
+-> knowledge-base MCP search
| |
| +-> Memos (quick notes)
| +-> Docmost (structured docs)
|
v
RAG-augmented response with citations
Key files
| File | Purpose |
|---|---|
tools/ai/dispatch_to_agent.sh |
Agent dispatch (Tower domains: docs, commercial, research) |
tools/run_knowledge_base_mcp.sh |
Knowledge-base MCP server runner |
specs/knowledge_base_mcp_spec.md |
MCP server specification |
test/test_knowledge_base_mcp.py |
MCP server tests |
docs/commercial/*.md |
Commercial documents for RAG upload |
mcp.json |
MCP server configuration (includes knowledge-base entry) |