What is the difference between naive RAG and advanced RAG?
- 1 min — naive pipeline as the baseline
- 2 min — pre / during / post retrieval menu
- 2 min — error analysis: three buckets → three fixes
- 1 min — latency cost of each addition
Naive RAG: fixed-size chunks → embed → top-k cosine → stuff into prompt. Advanced RAG adds work at three points. Pre-retrieval: query rewriting, decomposition, HyDE, metadata extraction, better chunking. Retrieval: hybrid dense+sparse, filtering, multi-index routing. Post-retrieval: reranking, compression, deduplication, ordering, groundedness verification.
Naive RAG is the right thing to ship first — it's your baseline. The advanced techniques are each a response to a specific observed failure, and adding them speculatively costs latency without proven gain.
"Advanced RAG isn't a tier you upgrade to, it's a menu you order from based on your error analysis." Then describe how you'd do that error analysis: sample 100 failures, label whether retrieval found the doc at all, whether it ranked it, and whether the model used it. Those three buckets point at three different fixes.
Org scope: keep the error-analysis labels in a shared tool so search, ML and product teams argue from the same 100 failures. Each 'advanced' technique ships behind a flag with an owner and an eval delta, so you can rip out ones that stopped earning their latency.
What the answer sounds like at each level
Naive RAG is embed, search, stuff; advanced adds things like reranking and query rewriting.
Organises techniques into pre-retrieval, retrieval and post-retrieval and ties each to a failure it fixes.
Starts from error analysis buckets (not found / not ranked / not used), adds techniques only on evidence, and accounts for each one's latency and maintenance cost.
Follow-ups the interviewer will push
1Error analysis shows 60% of failures are 'doc not in top-50'. What next?
That is a recall problem, so a reranker won't help. Add hybrid BM25 for exact terms, fix chunking or parsing, and add query rewriting for conversational queries. Re-measure Recall@50 on the golden set.
2Now latency budget is cut in half.
Drop the techniques with the worst quality-per-millisecond first: HyDE and multi-query expansion (each an LLM call) go before the reranker, which usually buys the most nDCG per 100 ms.
Red flags interviewers listen for
- Adds every advanced technique by default.
- Cannot say which failure a given technique addresses.
- No baseline before adding complexity.
How it's scored
| Dimension | What strong looks like |
|---|---|
| Structure | Groups techniques by pipeline stage. |
| Evidence-driven | Ties every addition to a measured failure bucket. |
| Cost awareness | Prices each technique in latency and LLM calls. |
Sources
- Precise Zero-Shot Dense Retrieval without Relevance Labels (HyDE) — Gao et al., 2022paper
- Reciprocal Rank Fusion outperforms Condorcet and individual rank learning methods — Cormack et al., 2009paper