PostgreSQL Semantic Search

Quick Start

1. Setup

CREATE EXTENSION IF NOT EXISTS vector;

CREATE TABLE documents (
    id SERIAL PRIMARY KEY,
    content TEXT NOT NULL,
    embedding vector(1536)  -- 1536-dim embedding
    -- Or: embedding halfvec(3072)  -- 3072-dim embedding (halfvec = 50% memory)
);

2. Basic Semantic Search

SELECT id, content, 1 - (embedding <=> query_vec) AS similarity
FROM documents
ORDER BY embedding <=> query_vec
LIMIT 10;

3. Add Index (> 10k documents)

CREATE INDEX ON documents USING hnsw (embedding vector_cosine_ops);

Docker Quick Start

# pgvector with PostgreSQL 17
docker run -d --name pgvector-db \
  -e POSTGRES_PASSWORD=postgres \
  -p 5432:5432 \
  pgvector/pgvector:pg17

# Or PostgreSQL 18
docker run -d --name pgvector-db \
  -e POSTGRES_PASSWORD=postgres \
  -p 5432:5432 \
  pgvector/pgvector:pg18

# ParadeDB (includes pgvector + pg_search + BM25)
docker run -d --name paradedb \
  -e POSTGRES_PASSWORD=postgres \
  -p 5432:5432 \
  paradedb/paradedb:latest  # `latest` is convenient for quick-start; pin to e.g. paradedb/paradedb:pg17 for reproducible builds

Connect: psql postgresql://postgres:postgres@localhost:5432/postgres

Cheat Sheet

Common Queries

-- Top 10 similar (cosine)
SELECT * FROM docs ORDER BY embedding <=> $1 LIMIT 10;

-- With similarity score
SELECT *, 1 - (embedding <=> $1) AS similarity FROM docs ORDER BY embedding <=> $1 LIMIT 10;

-- With a distance threshold — put the filter OUTSIDE a materialized CTE.
-- Filtering inline (WHERE (embedding <=> $1) < 0.3 ORDER BY ... LIMIT 10) makes
-- the executor apply the filter before the index returns LIMIT rows, so you get
-- fewer results than expected. pgvector documents this CTE form as the fix.
WITH nearest AS MATERIALIZED (
  SELECT id, content, embedding <=> $1 AS distance FROM docs
  ORDER BY distance LIMIT 10
) SELECT * FROM nearest WHERE distance < 0.3 ORDER BY distance;

-- Preload index (run on startup)
SELECT 1 FROM docs ORDER BY embedding <=> $1 LIMIT 1;

Index Quick Reference

-- HNSW (recommended)
CREATE INDEX ON docs USING hnsw (embedding vector_cosine_ops);

-- With tuning
CREATE INDEX ON docs USING hnsw (embedding vector_cosine_ops)
WITH (m = 24, ef_construction = 200);

-- Query-time recall. Set this: pgvector's default of 40 costs recall silently
-- (measured ~1.1 pp at 54k vectors for no latency saving; on a 22k-vector
-- corpus recall@20 went 96.2% -> 99.2% for +0.7 ms median). See indexing.md.
-- Query-time settings are connection-local. Use SET LOCAL inside a transaction
-- when a transaction pooler can hand each request a different connection.
SET hnsw.ef_search = 100;

-- Iterative scan for filtered queries (pgvector 0.8+; OFF by default)
SET hnsw.iterative_scan = relaxed_order;    -- or strict_order
SET ivfflat.iterative_scan = relaxed_order; -- IVFFlat has no strict_order

Decision Trees

Choose Search Method

Query type?
├─ Conceptual/meaning-based → Pure vector search
├─ Exact terms/names → Pure keyword search (FTS)
├─ Fuzzy/typo-tolerant → pg_trgm trigram similarity
├─ Autocomplete/prefix → pg_trgm + prefix index
├─ Substring (LIKE/ILIKE) → pg_trgm GIN index
└─ Mixed/unknown → Hybrid search
    ├─ Simple setup → FTS + RRF (no extra extensions)
    ├─ Better ranking → BM25 + RRF (pg_search extension)
    └─ Full-featured → ParadeDB (Elasticsearch alternative)

Baseline hybrid against vector-only before shipping it. Hybrid is the right default for mixed queries, not an automatic win: on one measured corpus pure vector beat a well-weighted hybrid on both recall and MRR and was 9× faster, while a badly weighted one lost 10 pp. The keyword arm still earns its keep for exact identifiers, which a needle-in-haystack eval set cannot see — so keep it, and judge it on queries that need it. Numbers in [hybrid-search.md](references/hybrid-search.md#hybrid-does-not-automatically-beat-pure-vector).

Choose Index Type

Document count?
├─ < 10,000 → No index needed
├─ 10k - 1M → HNSW (best recall)
└─ > 1M → IVFFlat (less memory) or HNSW

Choose Vector Type

Choose by dimensions, not by provider — the column type only depends on embedding size and pgvector's HNSW index limits.

Embedding dimensions (N)?
├─ N ≤ 2000  → vector(N)   — HNSW indexable directly
├─ 2000 < N ≤ 4000 → halfvec(N) — vector(N)'s HNSW limit is 2000; halfvec extends to 4000
└─ N > 4000  → vector(N) without HNSW, or quantize via dimensionality reduction

Common embedding dimensions are 1536 and 3072, but sizes vary by provider and model — check the provider's docs for the embedding you're using.

For multilingual / non-English content, prefer multilingual-tuned embedding models (look for "multilingual" in the model name). Models tuned only on English may handle compound words and inflection poorly.

Storage vs. index trick for 2000 < N ≤ 4000: keep the column as vector(N) (full float4, useful for future re-embedding or re-ranking experiments) and only cast at index creation and query time. This preserves precision on disk while staying within HNSW's dimension limit.

CREATE INDEX ON docs USING hnsw ((embedding::halfvec(3072)) halfvec_cosine_ops);
-- Query must cast identically so the planner picks the index:
SELECT * FROM docs ORDER BY embedding::halfvec(3072) <=> $1 LIMIT 10;

If storage is tight or you never plan to re-embed, use halfvec(N) as the column type directly.

Measure before adopting

Every optimization in this skill (hybrid fusion, reranking, query expansion, embedding-model swaps) can regress on a specific corpus. Vendor and paper benchmarks are usually English, general-domain, and their ordering does not reliably transfer. Real counter-examples, each measured rather than argued:

• Query expansion (HyDE) regressing Hit@5 by tens of points on a domain corpus.

On another, it found +1.1 pp more and ranked worse (MRR 0.557 → 0.536) at 6× the hybrid latency — a reranking loss dressed as a recall win.

• A widely recommended reranker regressing Hit@5 double-digits on multilingual text.
• Translating an off-language query into a keyword list rather than a

sentence: it helped the arm it targeted and finished 14 pp below doing nothing at all.

• Raising top-k from 15 to 30: recall +5.4 pp, and the share of generated claims

actually supported by a source fell 82.1 % → 78.8 %. Better retrieval, worse answer.

• The cheapest open embedding model beating the paid one on the target corpus,

reversing the leaderboard order.

Rule: build a domain eval set ([evaluation.md](references/evaluation.md)), then A/B each change. Adopt with ≥ +3 pp Hit@5 and p95 latency within budget; reject otherwise.

Three traps that make an A/B lie, each covered in [evaluation.md](references/evaluation.md#four-ways-a-measurement-lies-to-you): measuring one retrieval arm instead of the pipeline; treating retrieval metrics as the goal when an LLM consumes the results; and reading an offline sweep's absolute numbers as a production forecast. Use two eval sets — generated sentences and 1–3 word domain terms — because a change that helps one has measured as hurting the other.

Operators

SQL Functions

These are defined by this skill, not by pgvector. Install them by running the matching file from [scripts/](#scripts) — matchdocuments does not exist in a database that has not had semanticsearch.sql applied. Parameter names below are the real ones from the scripts — Supabase .rpc() binds by name, so a misspelled key fails at call time.

Semantic Search — scripts/semantic_search.sql

• matchdocuments(queryembedding, matchthreshold, matchcount) - Basic search
• matchdocumentsfiltered(queryembedding, filtermetadata, matchthreshold, matchcount) - With JSONB filter
• matchdocumentshalfvec(queryembedding halfvec(3072), matchthreshold, match_count) - halfvec column variant
• matchdocumentsdynamic(tablename, queryembedding, matchthreshold, matchcount) - Same search against any table name
• matchchunks(queryembedding, matchthreshold, matchcount) - Search document chunks

Fuzzy Search (pgtrgm) — scripts/fuzzysearch.sql

• fuzzysearchtrigram(querytext, similaritythreshold, max_results) - Trigram similarity search
• autocompletesearch(searchprefix, max_results) - Prefix + fuzzy autocomplete
• hybridsearchfuzzysemantic(querytext, queryembedding, maxresults, rrf_k) - Fuzzy + vector RRF
• weightedftssearch(querytext, ftslanguage, max_results) - FTS with title/content weighting

Hybrid Search (FTS) — scripts/hybridsearchfts.sql

• hybridsearchfts(queryembedding, querytext, matchcount, rrfk, fts_language) - FTS + RRF
• hybridsearchweighted(queryembedding, querytext, matchcount, semanticweight, keywordweight, ftslanguage) - Linear combination
• hybridsearchfallback(queryembedding, querytext, matchcount, rrfk, fts_language) - Graceful degradation (either input may be NULL)

These functions do not set hnsw.efsearch; the caller controls that query-time tradeoff. Set it on the active connection (or with SET LOCAL in the same transaction) before calling, or pgvector's default of 40 applies. Their keyword arm also wraps content in unaccent() and computes the tsvector per row; read the header of hybridsearch_fts.sql before using them on Finnish, Swedish, German or Turkish text, or on a table that has a tsvector GIN index.

Hybrid Search (BM25) — scripts/hybridsearchbm25.sql

• hybridsearchbm25(queryembedding, querytext, matchcount, rrfk) - BM25 + RRF
• hybridsearchbm25highlighted(queryembedding, querytext, matchcount, rrf_k) - With snippet highlighting
• hybridsearchchunksbm25(queryembedding, querytext, matchcount, rrf_k) - For RAG with chunks

Re-ranking (Optional)

Two-stage retrieval improves precision: fast recall → precise rerank with a cross-encoder. Use when results need higher precision and you have <50 candidates after initial retrieval.

Key rule: rerankers must be wrapped so a failure (missing key, HTTP error, timeout) returns null and the caller falls back to original retrieval order — never let a reranker outage break search.

For provider comparison, generic Promise<T | null> wrapper, and self-hosted options, see [reranking.md](references/reranking.md).

Multilingual / non-English content

Non-English corpora fail in specific, silent ways: the wrong FTS config skips stemming, unaccent merges distinct Finnish/Swedish/German words, zero-width characters glue onto tokens, every parser ANDs a long question into zero hits, and English-derived chunk caps overflow the embedding endpoint. The rules and fixes — FTS configs, prefix tsquery, synonym expansion, query translation, ParadeDB stemmer casts, per-language indexing, cross-language RRF fusion — are in [multilingual.md](references/multilingual.md). Read it before indexing anything that is not English prose.

References

• [fuzzy-search.md](references/fuzzy-search.md) - pg_trgm, fuzzy matching, LIKE/ILIKE, autocomplete, advanced FTS
• [paradedb.md](references/paradedb.md) - ParadeDB full-text search (Elasticsearch alternative)
• [vector-types.md](references/vector-types.md) - vector vs halfvec, dimensions, storage
• [indexing.md](references/indexing.md) - HNSW, IVFFlat, GIN parameters
• [hybrid-search.md](references/hybrid-search.md) - FTS, BM25, RRF algorithms
• [performance.md](references/performance.md) - Cold-start, memory, HNSW vs IVFFlat
• [evaluation.md](references/evaluation.md) - Eval-set construction, Hit@K / MRR, adoption thresholds, reranker/expansion benchmarking
• [reranking.md](references/reranking.md) - Two-stage retrieval, graceful fallback, when rerankers regress
• [multilingual.md](references/multilingual.md) - FTS configs and unaccent rules, invisible characters, prefix tsquery, query translation, per-language indexing, cross-language RRF

Scripts

• [setup.sql](scripts/setup.sql) - Extension and table setup
• [semanticsearch.sql](scripts/semanticsearch.sql) - Semantic search functions
• [hybridsearchfts.sql](scripts/hybridsearchfts.sql) - FTS hybrid functions
• [hybridsearchbm25.sql](scripts/hybridsearchbm25.sql) - BM25 hybrid functions
• [fuzzysearch.sql](scripts/fuzzysearch.sql) - pg_trgm fuzzy search, autocomplete, weighted FTS
• [indexes.sql](scripts/indexes.sql) - Index creation scripts
• [embeddings.ts](scripts/embeddings.ts) - Embedding generation helpers (TypeScript)

Common Patterns

TypeScript Integration (Supabase)

// Semantic search
const { data } = await supabase.rpc('match_documents', {
  query_embedding: embedding,
  match_threshold: 0.7,
  match_count: 10
});

// Hybrid search
const { data } = await supabase.rpc('hybrid_search_fts', {
  query_embedding: embedding,
  query_text: userQuery,
  match_count: 10,
  rrf_k: 60,
  fts_language: 'simple'
});

Drizzle ORM

import { sql } from 'drizzle-orm';

const results = await db.execute(sql`
  SELECT * FROM match_documents(
    ${embedding}::vector(1536),
    0.7,
    10
  )
`);

Troubleshooting

Compatibility

• pgvector: 0.8.6+ recommended as the safe floor (as of 2026-09). Feature history: 0.7.0 added halfvec/bit/sparsevec, 0.8.0 added iterative scans. Correctness history: 0.6.0–0.8.1 carry a parallel-HNSW-build buffer overflow (CVE-2026-3172 — leaks data from other relations or crashes the server), 0.8.2 fixed it, 0.8.3 fixed possible HNSW index corruption during vacuum, 0.8.4 fixed further HNSW vacuum errors, and 0.8.6 fixed an IVFFlat build integer wraparound on 32-bit systems (CVE-2026-18022). Verify current state in the CHANGELOG — the GitHub Releases tab is empty, releases ship as tags.
• pgsearch: Since 0.25.0 pgsearch depends on pgvector's vector type — install pgvector first. Check ParadeDB releases for latest.
• PostgreSQL: pgvector supports 13+; pg_search ships prebuilt binaries for 15+. Prefer the newest major your host offers.

Related Skills

For ParadeDB-specific questions, always apply the Documentation Fetch Policy in [references/paradedb.md](references/paradedb.md) — live docs at https://www.paradedb.com/docs/llms-full.txt are the authoritative source.

External Documentation

Core

• pgvector GitHub - Official extension, latest features
• PostgreSQL FTS - Built-in full-text search

Embedding providers

• OpenAI Embeddings - model list + dimensions
• Voyage Embeddings - includes multilingual model
• Cohere Embed - model list
• HuggingFace Hub - open-weight embeddings

Reranker providers

• Cohere Rerank
• Voyage Rerank
• Zerank
• Sentence Transformers - self-hosted cross-encoders

Hosting / extensions

• Supabase Vector Guide - Supabase-specific integration
• ParadeDB pg_search - BM25 extension documentation
• ParadeDB AI Docs - Fetch for latest ParadeDB API (always current)