Concepts

AI Search separates:

• Namespace — catalog of instances (parent resource; no per-namespace database)
• Instance — one SQLite authority plus immutable source objects on the configured Local or S3 backend
• env.AI — platform binding for Markdown Conversion only (toMarkdown / supported / aiGatewayLogId)

Indexing pipeline: upload → parse (document parser child) → chunk → embed (operator provider) → FTS + vector generations. Search modes: keyword, vector, and hybrid. Chat uses retrieval plus an optional generation provider; SSE streaming is supported when configured.

Providers are pinned by the operator (model revision, dimensions, tokenizer digest, timeouts). open-compute does not ship embedded model weights or claim Cloudflare-hosted model availability.

Not provided:

• Full Workers AI inference (run, models, AutoRAG, unrelated env.AI members)
• Cloudflare-hosted automatic provider selection, billing, dashboard connectors
• Global index placement / replication

Compatibility

Topic	Cloudflare	open-compute
Data path	Hosted AI Search	Per-instance SQLite + Local/S3 objects
Models	Workers AI	Operator OpenAI-compatible endpoints
Markdown Conversion	env.AI.toMarkdown	Same subset
Full Workers AI	Available	Not provided