pgvector Problems & Issues

Real user-reported problems, complaints, and limitations for pgvector. Based on analysis of 11 signals.

Overall score: 74/100

Issues by Category

performance4 issues

HNSW index build is very slow once the graph outgrows memory

The docs say builds are much faster while the graph fits in maintenance_work_mem and Postgres logs a notice when it no longer does. Users report long builds on big tables and need to raise memory and parallel workers.

Frequency: 8%Sentiment: -0.60/1004 sources
Impact
7

Filtered vector search returns fewer rows than LIMIT

Approximate indexes apply filters after the index scan. With ef_search 40 and a filter matching 10 percent you get about 4 rows. Iterative scans in 0.8 and partial indexes are the fixes.

Frequency: 8%Sentiment: -0.55/1005 sources
Impact
7

Recall drops and memory grows at tens of millions of vectors

Index memory competes with the rest of Postgres, so large workloads need bigger instances. Many teams then evaluate dedicated vector databases for scale.

Frequency: 6%Sentiment: -0.45/1003 sources
Impact
7

Queries do a sequential scan and ignore the vector index

The planner only uses the index for ORDER BY on a bare distance operator, ascending, with LIMIT, and with a matching operator class. Wrapped expressions silently fall back to a full scan.

Frequency: 8%Sentiment: -0.50/1004 sources
Impact
6
usability4 issues

Vectors over 2,000 dimensions cannot be indexed

Storage allows up to 16,000 dimensions but HNSW and IVFFlat on vector stop at 2,000, so 3,072-dim embeddings fail at CREATE INDEX. halfvec raises the index limit to 4,000.

Frequency: 8%Sentiment: -0.50/1004 sources
Impact
7

operator does not exist: vector <=> unknown or real[]

The distance operators need a vector on both sides. Untyped parameters or plain arrays from drivers and ORMs fail until you cast to vector or register the driver adapter.

Frequency: 8%Sentiment: -0.50/1004 sources
Impact
6

Dimension mismatch errors when you change embedding model

The column is fixed width, so inserting 768 values into a 1536 column fails. n8n, LlamaIndex and OpenAI forum users hit it and must re-embed or add a new column.

Frequency: 7%Sentiment: -0.50/1004 sources
Impact
6

IVFFlat gives poor results when built on little data

Docs advise building IVFFlat after loading data and sizing lists from row count. Built too early it returns weak recall until probes are raised.

Frequency: 5%Sentiment: -0.40/1003 sources
Impact
5
onboarding1 issue

pgvector error: extension "vector" is not available

The extension is named vector, not pgvector, and the server package must match your Postgres major version. Reports from Fly.io, Railway, DigitalOcean and Vultr threads show the control-file error after upgrades or on hosts that do not ship it.

Frequency: 9%Sentiment: -0.50/1004 sources
Impact
6
reliability1 issue

HNSW build fails on shared memory or runs out of RAM

Parallel builds in Docker need more than the default 64 MB /dev/shm, and a Cloud SQL thread reports failures on pgvector 0.8.0. Real OOM kills come from graph memory.

Frequency: 6%Sentiment: -0.60/1003 sources
Impact
7
integrations1 issue

Extension availability and version lag differ by managed host

Supabase and Neon let you enable it in a click or one SQL line; Railway and Vultr threads show templates without it. Neon supports the latest version plus one behind.

Frequency: 6%Sentiment: -0.35/1004 sources
Impact
5

Regret Patterns

Common reasons users regret choosing pgvector.

Embedding model changed after launch

The fixed-width column rejects new vectors, so you re-embed everything or add a new column.

Frequency: 7%Weeks after launch

Index built before sizing memory

HNSW build crawls or fails once the graph exceeds maintenance_work_mem.

Frequency: 7%First large data load

Filters hurt recall

Filtered queries return too few rows until iterative scans or partitioning are set up.

Frequency: 6%When metadata filters are added

3,072-dim embeddings cannot be indexed

Index creation fails above 2,000 dimensions for vector.

Frequency: 6%First index creation

Free tier too small for indexes

Small RAM and 500 MB or 1 GB storage limits fill quickly with embeddings.

Frequency: 5%Prototype to production

When pgvector Breaks Down

Specific scenarios where pgvector fails or underperforms.

Tens of millions of vectors on one instance

Index memory competes with the rest of Postgres and recall tuning gets harder.

Highly selective metadata filters

Few rows match so approximate scans return too little.

Index builds in a small Docker container

Default shared memory is too small for parallel HNSW builds.

Postgres major upgrade

Extension files for the old version are missing and vector.control is not found.

Free-tier project idle for a week

Supabase Free projects pause after 1 week of inactivity; Neon scales to zero after 5 minutes.