Vector stores

Where your embeddings live and how similarity search is performed.

---

MySQL (default)

• On MySQL ≥ 9, the plugin uses native VECTOR(n) columns and DISTANCE(..., 'COSINE') for fast cosine-distance search. A parse-only capability probe at activation avoids false positives.
• On older MySQL or MariaDB, it transparently falls back to storing embeddings as JSON LONGTEXT and scoring them in PHP with cosine similarity.

The Settings → Vector Database screen shows which mode is active.

JSON fallback limits

The JSON fallback loads up to 5,000 chunks per query and scores them in PHP. For larger knowledge bases, upgrade to MySQL 9 (native VECTOR) or use Cloudflare Vectorize.

Cloudflare Vectorize

A hosted vector index with a full REST v2 client: insert, query, delete-by-ids, delete-by-source-id, create-index, get-index. The text and metadata stay in your local chunks table; only the vectors live in Vectorize.

The Create / verify index button auto-detects the dimension from your embedding model, so you don’t have to look it up.

Choosing an engine

Option	Best for	Notes
Auto	Most users	Vectorize when configured, otherwise MySQL.
MySQL	Self-hosted, fully local	Native VECTOR on MySQL 9; JSON fallback otherwise.
Cloudflare Vectorize	Serverless scale	Hosted index; one Cloudflare account for everything.

For a fully offline stack, use MySQL + Ollama — no external requests are made. For a Cloudflare-only stack, Workers AI can serve both embeddings and chat, with Vectorize hosting the index.