RAG Chatbot Skill

Instructions

1. Setup

   • Create Python virtual environment
   • Install dependencies:

     pip install fastapi uvicorn pydantic openai qdrant-client asyncpg langchain-text-splitters tenacity python-dotenv
     

   • Create .env.example with: OPENAI_API_KEY, QDRANT_URL, QDRANT_API_KEY, NEON_DATABASE_URL, EMBED_MODEL, CHAT_MODEL

2. Data model

   • Neon tables:
     • documents(id, path, checksum, meta jsonb)
     • chunks(id, doc_id, content, meta jsonb)
     • sessions(id, user_id, prefs jsonb)
   • Qdrant collection book_chunks with payload: doc_path, module, week, tags, heading_path

3. Ingestion

   • Walk markdown glob, parse frontmatter
   • Split chunks (by headings and tokens)
   • Embed via OpenAI, upsert to Qdrant
   • Store doc/chunk metadata in Neon
   • Track checksum for incremental ingest

4. Query endpoints

   • /health - healthcheck
   • /ingest - trigger ingestion
   • /query - semantic search + LLM answer with sources
   • /query/selected - bypass vector search, inject user-selected text as context

5. Ops

   • Configure CORS for Docusaurus origin
   • Add logging with request IDs
   • Include safety: max tokens, fallback answers, latency budget

Examples

# Query endpoint structure
@app.post("/query")
async def query(request: QueryRequest):
    # 1. Embed query
    # 2. Search Qdrant
    # 3. Build context from chunks
    # 4. Call OpenAI with context
    # 5. Return answer + sources
    pass

# Run server
uvicorn main:app --reload --port 8000

Definition of Done

• FastAPI runs locally with env sample; healthcheck ok
• Ingestion populates Qdrant + Neon with at least sample doc; idempotent reruns succeed
• Query endpoints return grounded answers with cited headings
• Selected-text mode echoes user selection
• README snippet for running server and triggering ingest