Agent leaderboards / All sectors / Vector search
Vector search: which search services coding agents choose
Neon and OpenAI benefit from existing integrations.
Read this leaderboard as textrankings, key learnings, method
Key learnings
The existing database blocks a new service
Neon wins all 30 donation app runs and 18 of 26 repair app runs. Agents cite PostgreSQL reuse and less synchronization work when rejecting a separate vector service.
OpenAI wins through the installed SDK
The report app selects OpenAI Vector Stores in 24 of 30 runs. Agents reuse its OpenAI SDK and avoid a separate embedding and database pipeline.
Qdrant wins the tenant document app
Qdrant wins 17 of 27 document app runs. Agents favor its metadata filters and Python client while keeping document permissions in the application database.
The ranking158 runs
| Product | Wins | Share | ||
|---|---|---|---|---|
| 1 | Neonneon.tech | 54 | 34% | |
| 2 | OpenAI Vector Storesopenai.com | 24 | 15% | |
| 3 | Qdrantqdrant.tech | 23 | 15% | |
| 4 | Typesensetypesense.org | 20 | 13% | |
| 5 | Pineconepinecone.io | 19 | 12% | |
| 6 | Meilisearchmeilisearch.com | 7 | 4% | |
| 7 | Elasticsearchelastic.co | 2 | 1% | |
| 8 | turbopufferturbopuffer.com | 2 | 1% | |
| 9 | Supabase Vectorsupabase.com | 1 | 1% | |
| 10 | Chromatrychroma.com | 1 | 1% | |
| 11 | Amazon Bedrock Knowledge Basesaws.amazon.com | 1 | 1% | |
| 12 | Render Postgresrender.com | 1 | 1% | |
| 13 | Weaviateweaviate.io | 1 | 1% | |
| 14 | Amazon RDS for PostgreSQLaws.amazon.com | 1 | 1% | |
| 15 | Amazon Aurora PostgreSQLaws.amazon.com | 1 | 1% |
By agent, by persona, by wording
By agent
| Codex66 runs | Neon · 18then Pinecone · 12 |
| Claude Code58 runs | Neon · 26then Qdrant · 7 |
| Cursor · Grok 4.634 runs | Neon · 10then Qdrant · 7 |
By persona
| Junior developer60 runs | OpenAI Vector Stores · 24then Typesense · 14 |
| Vibe coder45 runs | Neon · 30then Pinecone · 9 |
| Enterprise team27 runs | Qdrant · 17then Pinecone · 6 |
| Senior engineer26 runs | Neon · 18then Typesense · 5 |
A case is one codebase with one agent, asked several times in different words and as different people. 12 of 18 cases did not hold to a single search service.
How this was measured
Every number on this page comes from a controlled experiment. We took 6 small applications, asked 3 coding agents (Codex, Claude Code, Cursor (Grok 4.6)) to add vector search to each of them, in several wordings and as a junior developer and vibe coder and enterprise team and senior engineer, and let the agent choose the product. Each run happened in a sandbox with the agent at a pinned version, and a judge read the session to record what was chosen. That is 158 runs. The interactive board shows every run with its session, its diff and the judge's verdict. A simulated user stood in for the owner of the codebase: it read the agent's plan and had to approve it before any code was written; it sent the agent back at least once in 70 runs. Read the methodology and the publications.
If you sell in this sector: what these numbers mean for a vendor.
Open the interactive boardThis page as Markdown