Hybrid Retrieval, Reranking, and Citations

Hybrid retrieval combines lexical search, which excels at exact terms and rare identifiers, with vector search, which captures semantic similarity. Candidate rankings can be fused and optionally reranked before bounded context reaches a model. Retrieval quality, answer groundedness, and citation correctness are separate measurements; a fluent answer does not prove any of them.

What will you be able to do?

Build PostgreSQL full-text candidates and vector candidates.
Fuse ranked lists without comparing incomparable raw scores.
Rerank and assemble bounded, diverse context.
Create citations that resolve to authorized source versions.

Prerequisites and mental model

Use a funnel:

text

authorized corpus → lexical/vector candidates → rank fusion → rerank
→ diversify/deduplicate → context budget → answer → citation validation

How can ranked lists be fused?

Reciprocal rank fusion uses rank positions instead of raw scores:

sql

WITH lexical AS (
  SELECT chunk_id,
         row_number() OVER (ORDER BY ts_rank_cd(search_vector, websearch_to_tsquery('english', $1)) DESC) AS r
  FROM signaldesk.knowledge_chunks
  WHERE organization_id = $2
    AND search_vector @@ websearch_to_tsquery('english', $1)
  LIMIT 50
), semantic AS (
  SELECT chunk_id,
         row_number() OVER (ORDER BY embedding <=> $3::vector) AS r
  FROM signaldesk.retrievable_chunks
  WHERE organization_id = $2
  ORDER BY embedding <=> $3::vector
  LIMIT 50
), fused AS (
  SELECT chunk_id, sum(1.0 / (60 + r)) AS score
  FROM (
    SELECT * FROM lexical
    UNION ALL
    SELECT * FROM semantic
  ) AS candidates
  GROUP BY chunk_id
)
SELECT chunk_id, score
FROM fused
ORDER BY score DESC, chunk_id
LIMIT 20;

The specific constant/candidate counts are evaluation parameters, not universal truth. The actual schema must ensure authorization in both branches.

Reranking may use a cross-encoder/model or deterministic features. Log candidate IDs and ranks, not sensitive content by default. Context construction should cap total tokens, avoid redundant adjacent chunks, preserve document/section structure, and separate instructions from retrieved text.

Why is this AI-native?

Hybrid retrieval turns PostgreSQL into a governed evidence service: relational policy and versions select eligible content, specialized indexes find candidates, evaluation chooses configuration, and citation lineage supports audit. The generator remains downstream and untrusted.

Prove it worked

Build labeled queries covering exact IDs, paraphrases, negation, stale versions, and no-answer cases. Measure lexical, vector, fused, and reranked recall@k/precision or ranking metric. Validate every generated citation resolves to a retrieved, authorized chunk/version and actually supports the claim.

Production judgment

Language dictionaries, stemming, tokenization, and multilingual content change lexical behavior.
Rerank calls add latency, cost, and another failure mode.
Raw score thresholds do not transfer cleanly across models/queries.
Context diversity can improve coverage but reduce nearest-score purity; evaluate.

Common mistakes

Exact product code missed: vector-only retrieval → add lexical path.
Scores added directly: incomparable scales → fuse ranks or calibrate on data.
Citation exists but unsupported: link presence mistaken for evidence → claim-level validation.

AI Pair-Programmer Prompt

text

Design a hybrid retrieval experiment. Require authorized corpus, lexical configuration,
vector model/metric/index, candidate counts, fusion method, reranker, dedup/diversity,
context budget, no-answer behavior, labeled queries, retrieval and answer metrics,
citation support validation, latency/cost, fallback, and versioned rollout.

Hands-on exercise

Create a 20-question dataset with exact codes and semantic paraphrases; compare three retrieval variants.

Acceptance criterion: the chosen variant wins on declared metrics within latency/cost budget and has zero unauthorized/stale citations.

Knowledge check

Why combine lexical and vector search?
Why fuse ranks rather than raw scores?
Does a resolvable citation prove support?

Answers

They have complementary exact-term and semantic strengths.
Raw scores have different scales/meanings.
No; the cited passage must actually support the claim.

Completion checklist

Candidate branches enforce identical authorization.
Fusion/reranking is versioned and evaluated.
Context is bounded and diversified deliberately.
Citations resolve and support claims.

Primary references

Hybrid Retrieval, Reranking, and Citations

How can ranked lists be fused?

Reciprocal rank fusion uses rank positions instead of raw scores:

sql

WITH lexical AS (
  SELECT chunk_id,
         row_number() OVER (ORDER BY ts_rank_cd(search_vector, websearch_to_tsquery('english', $1)) DESC) AS r
  FROM signaldesk.knowledge_chunks
  WHERE organization_id = $2
    AND search_vector @@ websearch_to_tsquery('english', $1)
  LIMIT 50
), semantic AS (
  SELECT chunk_id,
         row_number() OVER (ORDER BY embedding <=> $3::vector) AS r
  FROM signaldesk.retrievable_chunks
  WHERE organization_id = $2
  ORDER BY embedding <=> $3::vector
  LIMIT 50
), fused AS (
  SELECT chunk_id, sum(1.0 / (60 + r)) AS score
  FROM (
    SELECT * FROM lexical
    UNION ALL
    SELECT * FROM semantic
  ) AS candidates
  GROUP BY chunk_id
)
SELECT chunk_id, score
FROM fused
ORDER BY score DESC, chunk_id
LIMIT 20;

The specific constant/candidate counts are evaluation parameters, not universal truth. The actual schema must ensure authorization in both branches.

Prove it worked

Production judgment

Language dictionaries, stemming, tokenization, and multilingual content change lexical behavior.

Rerank calls add latency, cost, and another failure mode.

Raw score thresholds do not transfer cleanly across models/queries.

Context diversity can improve coverage but reduce nearest-score purity; evaluate.

AI Pair-Programmer Prompt

text

Design a hybrid retrieval experiment. Require authorized corpus, lexical configuration,
vector model/metric/index, candidate counts, fusion method, reranker, dedup/diversity,
context budget, no-answer behavior, labeled queries, retrieval and answer metrics,
citation support validation, latency/cost, fallback, and versioned rollout.