RAG retrieval pattern

Last updated Jul 17, 2026

Standard **embed → search → pack → generate** loop for question answering over your docs.

Pipeline

Documents → chunk → embed → vector store
Question  → embed → top-k similarity → prompt context → LLM → answer + cites

Index time (once per doc update)

| Step | Action |

|------|--------|

| Chunk | Split text (~300–800 tokens, overlap ~10–20%) |

| Embed | Same model as query time |

| Store | vector + raw text + metadata (source, id) |

Query time

| Step | Action |

|------|--------|

| Embed query | Same model/dimensions as index |

| Retrieve | Cosine similarity or vector DB search, k = 3–8 |

| Pack | Fit chunks in context window; label each chunk ID |

| Generate | Prompt: "Use only these notes…" + optional citations |

Minimal Python pattern

def rag_answer(question: str) -> str:
    q_vec = embed(question)
    top = vector_db.search(q_vec, k=5)
    context = "\n\n".join(f"[{c.id}] {c.text}" for c in top)
    return llm(f"Notes:\n{context}\n\nQ: {question}\nA:")

Failure modes

| Symptom | Fix |

|---------|-----|

| Wrong facts | Improve retrieval before changing model |

| No matches | Lower threshold; hybrid keyword search |

| Context overflow | Reduce k or summarize chunks |

Try the interactive lesson →

← All API docs