LarpDog

An agentic RAG chatbot that answers newcomer onboarding questions with cited sources, pulling from uploaded documents and GitHub repos. Pipeline: documents (PDF/Word/Slides/Markdown) + GitHub repo content → chunked & embedded → retrieved + reranked → LLM answer with citations back to repo/path or source doc. Stack: - API: FastAPI (api/app.py, ask.py, admin.py, repos.py, feedback.py) - Ingestion: docling-based parsing + chunking (ingest/) - Retrieval: ChromaDB (embedded, persistent) + BM25 hybrid search, bge-m3 embeddings, bge-reranker-v2-m3 reranking (index/) - Answering: LLM loop with citation enforcement (answer/loop.py, citations.py), logged to SQLite (answer_log.db) - LLM backend: configurable — DeepSeek (cloud) or local Ollama, switch driven - API: FastAPI (api/app.py, ask.py, admin.py, repos.py, feedback.py) - Ingestion: docling-based parsing + chunking (ingest/) - Retrieval: ChromaDB (embedded, persistent) + BM25 hybrid search, bge-m3 embeddings, bge-reranker-v2-m3 reranking (index/) - Answering: LLM loop with citation enforcement (answer/loop.py, citations.py), logged to SQLite (answer_log.db) - LLM backend: configurable — DeepSeek (cloud) or local Ollama, switch driven by content sensitivity - GitHub connector: MCP-based (connectors/github_mcp) syncing repos listed in repos.yaml - Frontend: simple HTML chat UI + password-gated admin views (frontend/) - Eval: golden-question harness (eval/golden.yaml, run.py, scorer.py) with an 80% pass bar including multi-hop questions - Deploy: Docker Compose, single-host pilot deployment (EC2 or on-prem), documented in deploy/, GO_LIVE.md