🤖 AI & LLMsOpenAIproduction

pgvector cosine similarity returns irrelevant results for short search queries

Fresh5 months ago

Mar 14, 20260 views

Confidence Score63%

63%

Problem

Semantic search using pgvector returns irrelevant results when the query is 1–3 words. Short queries produce low-quality embeddings because the model has insufficient context to encode a meaningful semantic direction. A query like 'login error' returns documents about unrelated errors. Hybrid search combining vector similarity with keyword matching (pg_trgm or full-text search) with Reciprocal Rank Fusion significantly improves short query results.

Unverified for your environment

Select your OS to check compatibility.

Your OS

OS version

Product version

1 Fix

Canonical Fix

Moderate Confidence Fix

59% confidence63% success rate3 verificationsLast verified Mar 14, 2026

Implement hybrid search: combine pgvector with full-text search using Reciprocal Rank Fusion

Low Risk

Short query embeddings lack semantic direction. Combining vector similarity with keyword full-text search (pg_trgm or tsvector) and merging results with Reciprocal Rank Fusion (RRF) compensates for weak embeddings on short queries.

Trust Score

3 verifications

63% success

Enable pg_trgm extension

Run once in your database:

sql

CREATE EXTENSION IF NOT EXISTS pg_trgm;
CREATE INDEX IF NOT EXISTS idx_issues_title_trgm ON issues USING gin(title gin_trgm_ops);

Run vector search and keyword search in parallel

Fetch top-K results from each:

typescript

const [vectorResults, keywordResults] = await Promise.all([
  prisma.$queryRaw`
    SELECT id, title, embedding <=> ${embedding}::vector AS dist
    FROM issues ORDER BY dist LIMIT 20
  `,
  prisma.$queryRaw`
    SELECT id, title, similarity(title, ${query}) AS sim
    FROM issues WHERE title % ${query} ORDER BY sim DESC LIMIT 20
  `,
])

Merge with Reciprocal Rank Fusion

RRF score = 1/(k + rank), summed across both lists:

typescript

function rrfMerge(lists: {id: string}[][], k = 60) {
  const scores = new Map<string, number>()
  for (const list of lists) {
    list.forEach((item, rank) => {
      scores.set(item.id, (scores.get(item.id) ?? 0) + 1 / (k + rank + 1))
    })
  }
  return [...scores.entries()].sort((a, b) => b[1] - a[1]).map(([id]) => id)
}

Validation

Search for 1–3 word queries. Results should include highly relevant documents that pure vector search missed.

Verification Summary

Worked: 3

Partial: 2

Failed: 3

Last verified Mar 14, 2026

Environment

Product: pgvector + OpenAI Embeddings
Environment: production

Submitted by

Alex Chen

2450 rep