Multi-Agent Research Platform

An orchestrated multi-agent workflow that researches EU-compliance topics over the Project 1 corpus and produces a cited brief — with a corrective-RAG loop.

• Package: multi_agent_research_platform · Python: 3.12
• Orchestration: LangGraph StateGraph (Planner → Researcher → Critic → Writer)
• LLM: OpenAI-compatible (HF router → Qwen2.5-72B-Instruct)
• Retrieval: Atlas $vectorSearch over Project 1's chunk corpus
• Embeddings: Voyage voyage-context-4

Workflow

START → plan → research → critic ──(insufficient)──▶ research
                              \──(sufficient)──────▶ write → END

Node	Role
plan	decomposes the topic into 3 focused sub-questions
research	retrieves evidence (vector search) per sub-question, de-duplicated
critic	judges sufficiency; if weak, proposes follow-up questions and loops back (corrective RAG, bounded by MAX_REVISIONS)
write	composes a grounded brief citing evidence inline as [n]

Setup

cp .env.example .env     # MONGODB_URI, VOYAGE_API_KEY, HF_TOKEN
make install

Build Project 1's corpus first — it is the knowledge base this platform researches.

Run

make run
curl -s localhost:8000/research -H 'content-type: application/json' \
  -d '{"topic":"obligations for high-risk AI systems that process personal data"}'

Returns {topic, sub_questions, brief, citations, revisions}.

Quality gate

make check     # ruff + mypy --strict + pytest (incl. corrective-loop test)

Generated from ai-portfolio-template.