Every AI tool gives you one confident answer. The problem? Science doesn't work that way.

"3,118 tokens. 80% agent agreement. 65% confidence. That's the difference between one opinion and a deliberated verdict."

Query: "Are there contradictions in BRCA1's role in triple-negative breast cancer?"
Result: 65% confidence · 4 contradictions found · 6 key findings · 8 methodology flags

Ingest papers from PubMed, arXiv, Semantic Scholar, or upload PDFs directly. Research Council extracts entities (genes, drugs, diseases, pathways) and maps their relationships into a Neo4j knowledge graph.

Type any biomedical research question. The system retrieves the most relevant subgraph and paper chunks — not 20,000 tokens of noise, but a focused ~2,000 token context.

Four specialized agents analyze in parallel:

Then they cross-review each other (12 peer evaluations). The Chairman synthesizes a final verdict.

Every answer includes:

• Confidence score — calibrated, not inflated
• Key findings — cited to specific papers
• Contradictions found — highlighted explicitly in red
• Methodology notes — quality flags on the evidence

Every conclusion is written back into the graph as a new node — with provenance trails connecting it to the source papers. The graph gets smarter with every query.

🧬 Contradiction Detection

"Are there contradictions in KRAS G12C inhibitor resistance mechanisms?"
Find where the literature conflicts before you design your next experiment.

💊 Hypothesis Validation

"Has metformin been tested in glioblastoma? What failed and why?"
Check if your hypothesis has been tried before spending $500K in the lab.

📚 Living Systematic Review

"Summarize recent evidence on GLP-1 agonists and neurodegeneration"
Get a weekly auto-updated synthesis as new papers are ingested.

Researcher Query
      │
      ▼
GraphRAG Layer (Neo4j + ChromaDB)
  Entities: Gene · Drug · Disease · Protein · Pathway
  Relations: CONTRADICTS · SUPPORTS · MENTIONS · STUDIES
      │
      ▼
LangGraph Orchestrator
  1. BigTool Agent — selects 2-4 relevant tools (not all 50+)
  2. Hybrid Retrieval — vector search + graph expansion
  3. Context Assembly — ~2,000 tokens (not 20,000)
      │
      ▼
LLM Council (Groq + OpenRouter)
  Stage 1: 4 agents analyze in parallel
  Stage 2: 12 cross-reviews
  Stage 3: Chairman synthesizes + scores
      │
      ▼
Answer + Writeback to Graph
  Confidence score · Citations · Contradictions · Methodology notes

• Docker (for Neo4j)
• Python 3.11+
• Node.js 18+
• Groq API key (free tier available)
• OpenRouter API key

git clone https://github.com/al1-nasir/research-council
cd research-council
cp .env.example .env

Edit .env:

GROQ_API_KEY=your_groq_key
OPENROUTER_API_KEY=your_openrouter_key
NEO4J_URI=bolt://localhost:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=password

docker run -d --name neo4j \
  -p 7474:7474 -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/password \
  neo4j:5-community

# Install uv if you don't have it
pip install uv

uv sync
uvicorn api.main:app --reload --port 8000

cd frontend
npm install
npm run dev

Open http://localhost:5173

Click Search Papers → search for any topic → select papers → click Ingest Selected

Then ask your first research question. Try:

"Are there contradictions in BRCA1's role in triple-negative breast cancer?"

✅ Runs on a standard laptop. No GPU needed. Embeddings use all-MiniLM-L6-v2 (80MB, CPU-only).

Research Council uses dynamic tool loading via langgraph-bigtool — instead of loading all tool definitions upfront (which wastes tokens), the agent searches for and loads only 2-4 relevant tools per query.

Traditional approach:  50 tools × 500 tokens = 25,000 tokens before doing anything
Research Council:      2-4 tools loaded dynamically = ~200 tokens

Real query cost:       3,118 tokens for a full 4-agent deliberation
Estimated cost:        ~$0.002 per deep research query

• 4-agent council with cross-review
• GraphRAG with Neo4j
• Contradiction detection
• Confidence scoring
• Multi-source paper ingestion
• Knowledge graph visualization
• Community detection (Louvain clustering)
• Temporal analysis (track understanding over time)
• Export council output as PDF
• MCP server integrations (V2)
• HuggingFace Space demo
• Custom agent roles

Contributions are very welcome. This is a V1 — there's a lot of room to grow.

Easy first issues:

• Add support for a new paper source (bioRxiv, ChemRxiv, Europe PMC)
• Add a new agent role (Statistical Agent, Ethics Agent, Replication Agent)
• Improve entity extraction prompts for specific domains
• Add Docker Compose for one-command setup
• Write tests for graph_tools.py or council/agents.py

• karpathy/llm-council — the multi-model council concept
• microsoft/graphrag — GraphRAG methodology
• langchain-ai/langgraph-bigtool — dynamic tool loading

MIT — use it, fork it, build on it.