Search and Ask AI

Every Readsmith site with a database gets search; add a model key and it gets semantic search and Ask AI, cited answers generated from your own content, on your own keys. Nothing routes through a third party you did not choose.

Configure

docs.yamlyaml

ai:
  chat:
    provider: anthropic
    model: claude-haiku-4-5-20251001
  embedding:
    provider: openai
    model: text-embedding-3-small

Providers: openai, anthropic, google, or gateway for chat; openai, google, or gateway for embeddings. Mix freely. Keys never go in config; they come from the environment, role overrides first, provider-native variables second.

Index the content

pnpm ai:index

Indexing is incremental by content hash: unchanged chunks are never re-embedded, so a typo fix costs one chunk, not a re-index. Run it after deploys (the compose image runs migrations on boot; indexing is the one manual step in this release).

The degradation ladder

Features turn off one at a time, never all at once:

Missing	What still works
Chat key	Everything except Ask AI
Embedding key	Ask AI and keyword search; the semantic arm is off
Provider down at request time	Search answers `degraded: true` with keyword results, and the UI says so
Database	The whole docs site, agent outputs included; search and Ask AI are off

Clients can read the current state from the capabilities endpoint rather than guessing.

Grounding and citations

Ask AI retrieves full chunks, never display previews, and answers cite the pages they came from. The same retrieval serves the on-site console and the MCP search_docs tool, so an agent and a reader get the same floor of quality.

Rate limits

On by default: Ask AI 10 requests per minute, search 60, per client IP. Tune in config or override at deploy time:

docs.yamlyaml

ai:
  limits:
    ask: { limit: 20, windowMs: 60000 }

Ask AI spends your model credits; the tight default is intentional. The security page covers the environment overrides and proxy IP handling.