Search indexed memory and return the passages that match. Prefer a specific question over keywords. Retrieving more than you need and picking among them yourself works better than trusting the first result: the ranking is vector similarity, which cannot tell that a passage merely repeats the question. Leave rerank off unless this service has a rerank profile configured — the default setup has none, and asking for it fails with RERANK_NOT_CONFIGURED.

Read a document's text around a character range, to see what a search result was cut off from. Pass the document_id and the source.start_offset / source.end_offset of a passage, widened by a few hundred characters on each side. Offsets past either end are clamped rather than refused, and total_codepoints tells a clamped range from an exhausted one.

List what is indexed, so a search can be narrowed or a gap in the memory can be recognised rather than guessed at.

Report which knowledge base this server is bound to and whether the service can answer searches right now.