Findings / run 2026-11-02
For RAG, the agents reach for Postgres first.
In 30 AI coding answers that build retrieval or semantic search, pgvector was installed 10 times, Qdrant and Chroma 5 each, Pinecone 2. Gemini never picked pgvector.
| Product | Installed | Claude | GPT | Gemini | Mentioned |
|---|---|---|---|---|---|
| pgvector | 10 | 4 | 6 | 0 | 16 |
| Qdrant | 5 | 0 | 3 | 2 | 9 |
| Chroma | 5 | 2 | 1 | 2 | 5 |
| FAISS | 2 | 1 | 0 | 1 | 3 |
| Pinecone | 2 | 0 | 0 | 2 | 3 |
Five tasks that need embeddings stored and searched (RAG over docs, semantic search, “similar items”), twice on each of Claude, GPT and Gemini. pgvector was installed in 10 of 30 answers, Qdrant and Chroma in 5 each, Pinecone and FAISS in 2 each, and 6 answers kept vectors in memory.
By model:
- GPT: pgvector 6 of 10. If you already have Postgres, GPT keeps you there.
- Claude: pgvector 4, Chroma 2, FAISS 1, in memory 3.
- Gemini: pgvector 0. Its picks were spread over Qdrant, Chroma and Pinecone (2 each), FAISS (1) and in-memory (3). Both of Pinecone's installs came from Gemini.
Qdrant was named in 9 answers and installed in 5, the best hit rate of the dedicated vector databases.
What it means
The default for “add RAG to my app” is now “add a column to the database you already have”. Dedicated vector databases win on specific prompts: Chroma for a Python “RAG over my PDFs” script (5 of 6 answers), Qdrant for “5 million embeddings” (4 of 6), and on Gemini in general. Samples are small (10 per model), so the per-model splits are a direction, not a verdict.
Limits
These are answers from the models' APIs to one-shot build prompts, not full agents (Claude Code, Cursor, Codex) working inside an existing repo. Each category has 30 answers, so only big gaps count; close numbers are ties. How we measure, and every answer is in the raw data.
See the data
- Vector database: all prompts and answers
- LLM observability: all prompts and answers
- Search: all prompts and answers
- Postgres hosting: all prompts and answers
More findings
- Supabase is mentioned, never installed: the Postgres host AI agents pick
- Drizzle vs Prisma: which ORM Claude, GPT and Gemini install
- Sentry and Playwright: the only tools AI agents install every time
- Which email API do AI agents install? Resend 16, Postmark 4, Mailgun 0
- Radix UI is named 11 times and never installed: shadcn/ui takes the pick
- Datadog was named 9 times by AI coding agents and installed 0 times
- Cloudflare R2 vs S3: AI agents name R2, then install the AWS SDK
- Where AI coding agents deploy: Vercel 12, Netlify, Railway, Render and Fly.io 0
Did we get something wrong?
A tool we missed, a pick that looks wrong, a prompt we should add? Tell us. We read every note and fix the data in the next run.
Is your product losing these picks?
Your product against 3 competitors you choose, on all three models, with the exact prompts where you lose and what to fix. $19, delivered within 24 hours.