RAG Over Your Docs

Answers questions grounded in a document set instead of the model's memory. Chunks source docs, embeds them, retrieves the top-k relevant chunks per query, and asks the model to answer only from retrieved context, with inline citations back to the source.

1. Ingest — split docs into ~500-token chunks with ~50-token overlap.
2. Embed — store vectors + source metadata in a vector DB.
3. Retrieve — embed the query, pull top-k (start with k=5) by cosine similarity.
4. Answer — prompt the model with the chunks and require citations.
5. Fallback — if no chunk clears a similarity threshold, say 'not found in the docs' rather than guessing.

Answer the question using ONLY the context below.
Cite each claim as [source: <file>#<chunk>].
If the context doesn't contain the answer, say so.

Context:
{retrieved_chunks}

Question: {user_question}

Require citations and reject answers without them. Set a minimum similarity threshold so off-topic questions return 'not found' instead of a hallucination. Log the retrieved chunk IDs with every answer so you can audit wrong results.