Index files or directories into a vector collection for semantic search. Safe to re-run: files whose content hash is unchanged are skipped. Honors .gitignore, skips binaries/oversized files, and uses language-aware chunking (markdown sections, code symbols, prose paragraphs).

Index arbitrary text (notes, snippets, memory entries, error logs, user-provided context) into a collection. Use a stable uri to replace previous content for the same uri, or pass replace:false to append a new version.

Run a semantic similarity query against indexed content. Returns ranked chunks with file path, line range, score (cosine similarity, 1 = identical), and text. Supports narrowing by path prefix, exact path, language, and content kind. Increase limit or use get_context when results look truncated.

Return contiguous chunks around a search hit (by chunk_id or document_id + ordinal) so you can read a whole function/section instead of an isolated fragment. Use after search when a result looks cut off.

List documents already indexed in a collection, with chunk counts and sizes. Use to check what is available before searching, or to confirm an index_paths run covered what you expected.

List all vector collections with their embedding model, dimensionality, document/chunk counts, and on-disk size.

Remove documents from a collection by document id, exact uri, or path prefix. Vectors and chunks are deleted together. This is irreversible and requires at least one selector.

Permanently delete an entire collection and all its vectors, chunks, and documents. Irreversible. Prefer delete_sources when removing only some documents.

Show database path, embedding model/provider, collection count, and totals. Useful for verifying the server is healthy and which model is active before diagnosing empty search results.

Return the full untruncated text of one chunk by its id, with its document metadata.