Give it a Git URL → it builds a code knowledge graph (who calls what, deterministically) + a semantic index (embeddings + BM25, RRF-fused, cross-encoder reranked) + a curated OKF knowledge bundle — then answers questions with an LLM that reasons over real relationships, streams tokens, and cites files & lines. Includes a 3D graph UI with precomputed layout and a confidence overlay.
question → rewrite → dense ×(variants×spaces) + BM25 → RRF → graph expand → rerank → LLM → answer + citations
ingest → clone → hash-diff → parse once → chunks+headers → embed (versioned) → graph build+reconcile → 3D layout → OKF
# 0) Redis
docker run -d -p 6379:6379 redis:7
# 1) API (terminal A)
cd backend && python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt && cp .env.example .env # add GEMINI_API_KEY for chat — optional: without it (or when Gemini is down) chat falls back to local Ollama (`ollama serve` + a pulled model)
uvicorn app.main:app --reload
# 2) Worker (terminal B)
cd backend && source .venv/bin/activate
python -m app.workers.ingestion_worker
# 3) Frontend (terminal C)
cd frontend && npm install && npm run dev # → localhost:5173Paste a repo URL, press Index repo, then chat / analyze / explore the 3D graph.
cd backend && pytest tests/ -v # unit tests + graph validation harness
python -m app.eval.run_eval ai-copilot-self # retrieval quality reportBeginner-oriented, file-by-file docs live in docs/ — including a complete rebuild checklist to reimplement the project from scratch. Strategy & plan: GraphRAG_Hybrid_Strategy.md (why) and AI_Copilot_Build_Handoff.md (what).
| Phase | Scope | Status |
|---|---|---|
| 0 Stabilize | chunker fix, JSON logging, eval harness | ✅ built |
| 1 Sharper RAG | code embedder + versioning, RRF, context headers, rewrite, caches | ✅ built |
| 2 The graph | store + two-pass builder, confidence edges, reconciliation, validation harness | ✅ built |
| 3 Fuse + see | graph-expanded retrieval, graph-aware prompts, 3D UI, OKF bundle | ✅ built |
| 4 Ship + scale | MCP server, webhook, HNSW, consumer groups/DLQ, precision tier | ▢ designed, not built |