Your search can return plausible matches while silently missing eligible records after a tenant, category or attribute filter. Before you change the index settings again, find out what your current retrieval path actually returns.

I compare your filtered pgvector HNSW search with exact PostgreSQL search, measure latency and confirm which index the database actually uses. You receive a diagnosis you can reproduce and use to decide what to change next.

A baseline report comparing exact results with your current approximate search: recall, result counts, latency and verified index use.

A bounded comparison of the current configuration and up to two agreed query-setting alternatives.

Examples of missed matches, short result sets or unexpected query plans where observed.

Reproducible commands and configuration, plus a written findings memo with prioritized recommendations and evidence limits.

One written clarification and factual-correction round.

$900 covers one PostgreSQL database, one vector table and embedding setup, and one HNSW retrieval path. We agree up to 100 total test queries across up to three filter patterns before measurements begin. The configuration comparison runs against a bounded dataset and workload agreed during intake.

Work uses staging, a suitable replica, or agreed outputs generated in your environment. Exact comparisons can be resource-intensive, so the run plan must fit the available infrastructure. No production changes are included.

I built pgvector-filterbench to compare filtered HNSW results against exact search and check recall, latency and actual index use. I also built pgvector-index-manager for planning, verifying and controlling HNSW index changes. These tools support the work; the deliverable is an evidence-based diagnosis of your system.

Measured accuracy here means recall against exact vector search. Human relevance, extraction quality and final AI-answer quality require separate evaluation.

Implementation, production deployment, new relevance labeling, embedding changes and chunking or retrieval rearchitecture are outside this fixed scope. If the findings justify further work, you receive a separately scoped recommendation.

Delivery: seven calendar days from confirmed scope, access and complete inputs.

I built the open-source tools behind this work:

pgvector-filterbench project · Source code

pgvector-index-manager project · Source code