📚 RAG (Retrieval-Augmented Generation)
Pipeline
flowchart LR A[Docs] --> B[Parse & clean] --> C[Chunk] --> D[Embed] --> E[(Vector + keyword index)] Q[User query] --> R[Rewrite / expand] --> S[Hybrid search] --> T[Rerank] --> U[Build prompt w/ citations] --> V[LLM] --> W[Answer + sources] E --> S
Checklist (core → advanced)
- Ingestion: parsing PDFs/HTML, metadata, incremental updates, deletions
- Chunking strategies: fixed, recursive, semantic, structure-aware; overlap; parent-child chunks
- Embedding models: choosing, dimensions, cost; re-embedding on model change
- Vector stores: pgvector (HNSW/IVFFlat indexes) ⭐, Qdrant, OpenSearch k-NN
- Hybrid search (BM25 + vector) with reciprocal rank fusion ⭐
- Reranking (cross-encoders / rerank APIs)
- Query transformation: rewriting, multi-query, HyDE, decomposition
- Metadata filtering and access control (per-user document permissions!)
- Contextual retrieval (adding context to chunks before embedding)
- Citations and “I don’t know” behavior
- Agentic RAG (the agent decides when and what to retrieve); GraphRAG (awareness)
- Evaluation: retrieval (recall@k, MRR, nDCG) and generation (faithfulness, answer relevance) → Evals, Guardrails & LLMOps
- Long context vs RAG: the trade-offs
🧪 Labs (🟢 warm-up → 🟡 core → 🔴 hard → ⚫ boss)
- 🟡 Ingestion pipeline (W18) → a baseline naive RAG with an eval set of 100 Qs
- 🔴 Hybrid + RRF + rerank + contextual chunks; measure each step’s gain
- 🔴 Document ACLs at query time + a leakage test
- ⚫ BM25 + HNSW from scratch in Go vs pgvector (recall/latency curves)
🧠 Cognitive tasks
- Error analysis on 30 failures → categorize → fix the biggest bucket
- Constraint flip: 1,000× more documents; what breaks first?
🛰️ Orbit integration
- UC2 knowledge copilot; RAG steps in UC3, UC4, UC6, UC7, UC8, UC10
Go deeper