Practical guide. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.
For: Developers building document and knowledge search
Before you start: A small corpus with stable document IDs · Questions with known supporting passages
⚡ TL;DR — the 30-second version
- Include exact IDs and paraphrases in the evaluation set.
- Measure retrieval separately from answer generation.
- Permission filters must apply before model context is assembled.
Semantic retrieval is useful for paraphrases. Lexical retrieval is useful for exact product codes, error messages and rare names. A hybrid pipeline combines those signals, but more machinery only helps if it improves relevance on the kinds of questions your users ask.
Start with a baseline and a labeled query set. Include both 'how do I reset this device?' and 'what does error E-1042 mean?' A single average score can hide a retrieval system that performs well on paraphrases and fails every exact identifier.
“An embedding can recognize a concept and still miss the one serial number the user needs.”
Field guide / 01
01 / Establish exact and lexical baselines
With pgvector, exact nearest-neighbor search is the default. Keep that as a reference when adding an approximate index. The official documentation describes HNSW and IVFFlat and warns that approximation changes the returned results.
For Qdrant, the reviewed hybrid guide uses dense and sparse queries with prefetch and fusion. Either way, hold the embedding model and corpus version fixed while testing retrieval parameters. Otherwise you cannot attribute the difference.
Field guide / 02
02 / Combine candidates, not incompatible raw scores
Scores from separate retrieval methods may live on different scales. Rank-based fusion can combine candidate lists without assuming a dense cosine score equals a lexical score. Reranking is a subsequent operation and should have its own latency and cost budget.
Measure recall at the candidate stage, then measure the quality of the final evidence set. If the supporting passage never becomes a candidate, the answer model cannot recover it reliably.
| Measure | Question it answers | Failure it reveals |
|---|---|---|
| Recall@k | Was supporting evidence retrieved? | Missing the required source |
| MRR | How early was the first relevant result? | Useful evidence buried too deep |
| Citation support | Does the answer's citation prove its claim? | Correct retrieval, unsupported synthesis |
| Filtered latency | How long under real access constraints? | A fast demo that slows down per tenant |
Field guide / 03
03 / Carry authorization through retrieval
Every chunk needs document identity, version and access metadata. A permission filter is not a UI preference: it belongs in the retrieval contract. Test two users with overlapping vocabulary and different document access, then verify the unauthorized passages never reach the model.
Selective metadata filters interact with approximate search performance and recall. Benchmark with the same tenant and category predicates your application uses, not only an unrestricted global query.
Field guide / 04
04 / Pick the simplest operational fit
If your data and access logic already live in PostgreSQL, pgvector can reduce data synchronization and operational overhead. A dedicated retrieval service can be useful when indexing and query pipelines need independent scaling or specialized operations.
The deciding evidence is your own relevance, latency, ingestion freshness and operations budget. Avoid selecting a database from a benchmark run on unrelated dimensions, filters or hardware.
Before you ship
0/5Use this as a review aid. Checks reset when you leave this page.
Sources & verification
Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.
- 01pgvector — official extension documentationpgvector↗
- 02Qdrant — hybrid semantic and lexical searchQdrant↗
- 03LlamaIndex — response and retrieval evaluationLlamaIndex↗
