Recursive branching truth engine. Builds research trees, not flat lists — each claim gets independently verified, each sub-topic gets its own branch, contradictions get investigated from both sides. The tree grows until convergence.
Each iteration: generate queries → search all providers → fetch → extract claims → triage (ACCEPT/VERIFY/DEEPEN/COUNTER) → check convergence → recurse into child branches. After all branches converge: cross-validate pending claims, synthesize final report.
 |
 |
Four Docker containers on a bridge network:
| Service | Port | Purpose |
|---|---|---|
| orchestrator | 8082 | FastAPI backend + React frontend + MCP server |
| vllm | 8000 | Local LLM inference (CUDA) |
| searxng | 8888 | Metasearch engine proxy |
| hiveminddb | 8100 | Knowledge graph + vector DB |
Two-model architecture:
- Bulk model (small/fast) — claim extraction, query generation, triage
- Synthesis model (larger) — cross-validation, final report
# Clone with submodule
git clone --recurse-submodules https://github.com/NodeNestor/Dendrite.git
cd Dendrite
# Configure (copy and edit)
cp .env.example .env # Set GPU_DEVICE, models, etc.
# Start everything
docker compose up -d
# Open web UI
open http://localhost:8082
# Or use the API
curl -X POST http://localhost:8082/api/research \
-H "Content-Type: application/json" \
-d '{"question": "Is fusion power viable by 2035?"}'All research parameters are configurable via .env, the API (PUT /api/config), or the web UI:
| Group | Parameters |
|---|---|
| Tree Structure | max_depth, max_branch_iterations, verification_iterations, queries_per_iteration |
| Search Width | urls_per_iteration, results_per_provider, max_concurrent_fetches, max_concurrent_llm |
| Verification | verification_threshold, min_independent_sources, max_concurrent_verifications, verification_fetch_count |
| Convergence | min_convergence_iterations, diminishing_returns_threshold, coverage_target |
Dendrite exposes 7 MCP tools for use with Claude Code or any MCP client:
| Tool | Purpose |
|---|---|
investigate |
Start a full research tree |
investigate_status |
Check progress of running investigation |
investigate_result |
Get final results + synthesis |
verify_claim |
Cross-validate a single claim |
search_knowledge |
Search past trees + HiveMindDB |
add_provider |
Register a custom data provider |
configure |
Update models, concurrency, depth limits |
# Run in MCP mode (stdio)
python -m orchestrator.src.main mcp- Branch types: investigation (main), verification (independent check), deepening (sub-question), counter (opposing evidence)
- Claim statuses: pending → verified / refuted / contested / accepted
- Convergence: min iterations, diminishing returns, zero new claims, LLM coverage target
- Cross-validation: 3-pass — rate claims → independent search → source independence check
- Providers: Modular ABC — plug in web, academic, local files, or any data source
- Docker + Docker Compose
- NVIDIA GPU with recent CUDA driver
- No external API keys required (all sources are free)
Private — NodeNestor