RAG Expert

RAG Engineering Expert

Garbage retrieval in, hallucination out. Answer quality is bounded by what you retrieve, so fix retrieval before touching the prompt. Measure retrieval and generation separately.

When to Use

• Building a RAG / "chat with your docs" / knowledge-assistant system.
• Tuning chunking, embeddings, retrieval, hybrid search, or reranking.
• Poor answers: missing context, wrong/irrelevant chunks, hallucinations, no citations.
• Choosing a vector store or designing metadata/filtering.

When NOT to Use

• Prompt wording/format only → prompt-engineering-expert.
• Training/fine-tuning a model → deep-learning-expert.
• Eval harness/regression design → llm-testing-expert.
• Serving infra/API shape → api-design-expert.

Core Principles

1. Ingestion & chunking

• Chunk on semantic boundaries (headings, paragraphs, code blocks), not fixed byte counts. Start ~256–512 tokens with ~10–15% overlap; tune per corpus.
• Preserve structure: keep tables/code intact, prepend section/title context to each chunk. Attach metadata (source, title, section, URL, timestamp, ACL) and a stable chunk ID for citations and filtering.

2. Retrieval quality (the highest-leverage area)

• Use a strong, task-matched embedding model; query and document embeddings must come from the same model/version. Re-index when you change models.
• Hybrid search: combine dense vectors with keyword/BM25 and fuse (e.g., RRF). Pure vector search misses exact terms — error codes, names, IDs, acronyms.
• Add a cross-encoder reranker over the top ~20 candidates and keep the top ~3–5. This is usually the biggest single quality win.
• Apply metadata filters (tenant, recency, doc type, ACL) — and enforce access control at retrieval, never just in the prompt.

3. Generation & grounding

• Pass only the reranked chunks that fit a sensible context budget (more is not better — it dilutes and raises cost/latency). Mark each with a source id.
• Instruct the model to answer only from context, cite sources, and say "I don't know" when context is insufficient. Return citations to the user for verifiability.

4. Evaluate & iterate

• Measure retrieval (recall@k, MRR, hit-rate) and generation (faithfulness/groundedness, answer relevance) separately on a labeled set. Low recall → fix chunking/retrieval/reranking; high recall but bad answers → fix the prompt/context budget.

Decision Guide

Common Mistakes

• Fixed-size character chunking that splits sentences/tables → semantic chunking + overlap.
• Mismatched query/doc embedding models → silently poor recall.
• Vector-only search → misses exact-match terms; go hybrid.
• No reranker → top-k is noisy; cross-encoder rerank.
• Stuffing 50 chunks into context → dilution, cost, lost-in-the-middle; retrieve more, send few.
• No citations / no "I don't know" path → unverifiable, hallucination-prone.
• ACL only in the prompt → data leakage; filter at retrieval.

Examples

Retrieve → hybrid → rerank → grounded prompt

dense = store.search(embed(query), top_k=20, filter={"tenant": tid, "acl": user.groups})
sparse = bm25.search(query, top_k=20)
candidates = rrf_fuse(dense, sparse)                 # reciprocal-rank fusion
ranked = reranker.rank(query, dedupe(candidates))[:5]  # cross-encoder

context = "\n\n".join(f"[{i+1}] {c.text}\n(source: {c.meta['title']})"
                      for i, c in enumerate(ranked))
prompt = (
    "Answer using ONLY the context. Cite sources as [n]. "
    "If the answer is not in the context, say you don't know.\n\n"
    f"Context:\n{context}\n\nQuestion: {query}"
)

See Also

• prompt-engineering-expert — structuring the grounded generation prompt.
• llm-testing-expert — measuring faithfulness and retrieval metrics.
• sql-expert — metadata filtering / pgvector storage.
• api-design-expert — streaming retrieval/answer endpoints.