Skip to content
⚠️ Before You Start
Risk: Low
Duration: Half a day to decide, 1-2 days with a benchmark
Last Validated: Not yet validated (draft)
Kind: Decision guide
Risk Note: Low (design decision; a poor choice shows up as weak relevance or extra cost, not an outage)

Choose the retrieval mode: keyword, vector, hybrid, semantic ranker or external re-ranker

Draft outline: not validated

Drafted with AI from official sources on 16 Oct 2026. Do not use on a live system until it is written, dry-run by someone other than the author and signed off (planned 2026-12-04). Author: Deeparani · dry-run: Gayathri Ramachandran · sign-off: Prasanna. Work items: RAG-08-T01, RAG-08-T02, RAG-08-T03, RAG-08-T04.

1. Summary

At design time for a new corpus, or when retrieval quality is poor and the team is debating whether to add a re-ranker.

Risk: Low (design decision; a poor choice shows up as weak relevance or extra cost, not an outage). Typical duration: Half a day to decide, 1-2 days with a benchmark.

2. When to use and when not to

Use it when: At design time for a new corpus, or when retrieval quality is poor and the team is debating whether to add a re-ranker.

Do not use it when: to be written by the author.

3. What you must know first

  • Hybrid search merges keyword and vector results with Reciprocal Rank Fusion (RRF)
  • Vector weighting multiplies a vector query's contribution to RRF; the default is 1.0
  • Semantic ranker re-scores only the top 50 results and reports @search.rerankerScore from 0 to 4
  • Semantic ranker is billed per query when the search string is not empty
  • An external re-ranker such as Cohere Rerank adds a network call and a per-call cost
  • Full-text recall in hybrid queries stops at 1,000 matches by default; maxTextRecallSize (preview) raises it

4. Options

Outline of the decision process; the author describes each option.

  1. Classify the corpus and queries: exact terms and codes, natural language, or mixed.
  2. Start from the default: hybrid (keyword plus vector) with semantic ranker.
  3. Benchmark keyword only, vector only, hybrid, and hybrid with semantic ranker on the labelled queries.
  4. Add an external re-ranker only if hybrid with semantic ranker misses the agreed target.
  5. Compare quality gain with added p95 latency and cost per 1,000 queries.
  6. Record the decision and the numbers in the decision table.
  7. Revisit when the corpus type, languages or query mix change.

Draft snippet: Hybrid query with a text vector query and semantic ranking (POST /indexes/{index}/docs/search)

{
  "search": "parental leave for contractors",
  "vectorQueries": [
    {
      "kind": "text",
      "text": "parental leave for contractors",
      "fields": "contentVector",
      "k": 50,
      "weight": 1.0
    }
  ],
  "queryType": "semantic",
  "semanticConfiguration": "default",
  "captions": "extractive",
  "select": "id, title, chunk",
  "top": 10
}

Snippets to add:

  • snippets/search/queries/hybrid-semantic.json: Hybrid query with semantic ranker (json)
  • snippets/decisions/retrieval-mode.md: Retrieval mode decision table (markdown)
  • snippets/eval/benchmark_modes.py: Benchmark the four modes (python)

5. Decision table

Situation Choose Why
to be written

To be written by the author from the outline above.

7. Known issues and fixes

Symptom Likely cause Fix
Exact part numbers or codes are missed Vector-only retrieval does not match rare tokens reliably Use hybrid so keyword matching catches exact terms, and give codes a field with a suitable analyser.
Semantic ranker gives no gain Semantic configuration lists fields in the wrong priority or very long fields Put title and key content first; input beyond the token limit is ignored.
Latency budget exceeded after adding a re-ranker Too many candidates re-ranked per query Re-rank a smaller candidate set and measure p95 again.

8. Monitoring and alerts

Signal Alert when Action
SearchLatency platform metric Above the agreed budget for 15 minutes after a ranking change Reduce candidate count or revert the change.

9. AWS and open-source equivalents

Amazon OpenSearch Service supports hybrid search through a normalisation processor in a search pipeline, and Amazon Bedrock offers a Rerank API. With pgvector, combine PostgreSQL full-text and vector results with RRF in SQL.

10. Validation record

Pending. Planned dry-run scenario: A second engineer runs the four modes on a sandbox index with the labelled set. Pass: the decision table holds recall@10 and nDCG@10, p95 latency and cost per 1,000 queries for every mode, and the chosen mode is justified by those numbers.

11. Sources