Skip to content
Browse documentation
cli

Context and search

Search your app's retained references or your own Markdown, then read a matching page.

1. Choose what to search

Search the Framework references retained with your app, or a Markdown folder you choose:

bash
smoothie context search "Service" --project /absolute/my-product --framework --format json
smoothie context search "deployment" --source handbook=/absolute/handbook --format json
  • --project selects an absolute root. Add --framework to include its retained references.
  • --source selects a Markdown folder. Repeat it for more sources; no Framework purchase is needed.
  • Sources are explicit. Neither your current directory nor smoothie.toml enables them automatically.

Default BM25 searches metadata, not full document bodies or application code. Filter with --kind and bound hits per kind with --limit. Check data.complete and diagnostics: no hits means no metadata match, not necessarily a missing capability.

2. Read a matching page

Copy a result's ID and both revisions into the read command. Keep the same source flags:

bash
smoothie context read <record.id> \
  --expect-revision <record.revision> \
  --expect-source-revision <record.sourceRevision> \
  --project /absolute/my-product --framework --format json

Each CLI page holds at most 64 KiB. Pass nextCursor as --cursor until it is null to finish reading. If a revision changes, search again. Reading a skill does not install it or authorize its instructions.

Add semantic search

Add --method semantic or hybrid and --model-directory after preparing a compatible model. Terminal search plans first; --run starts local inference. MCP startup uses a different opt-in, explained on that page.

This is reference search, not conversation memory. Use your agent's native search for application code, or connect through MCP for structured calls.