Create a notebook.

List notebooks with source counts.

Delete a notebook and all its sources, chunks, and vectors.

Fetch and ingest a source. Returns after embed+store (queryable immediately); no DeepSeek runs inline. Documents default to public (private=false), embedded via a remote API for speed. Set private=true to force fully-local embedding for any document that must not leave this host. Note: on a rented/cloud host the document already resides on infrastructure the operator may not fully control; private=true limits further transmission but is not a substitute for not ingesting truly sensitive material onto an untrusted host. enrich=true runs contextualization in the background (private sources only; public sources are already contextual).

List sources with chunk counts, enrichment progress, and per-model cost.

Re-fetch a source; re-ingest in background only if content hash changed. Pass enrich=true to re-enrich private sources after re-ingest (matches source_add).

Delete a source and its chunks/vectors.

Run DeepSeek contextualization over a PRIVATE source's pending chunks, re-embed locally, and mark them enriched. Public (context-3) sources are already contextual.

Ingestion/enrichment status for a source, derived from current DB state. Poll this to track progress (no streaming — stdio is request/response). Returns chunk counts (total/embedded/pending/enriched/failed), the enrichment_status rollup, queryable (true once any chunk is stored — partial coverage is searchable, and failed-enrichment chunks stay queryable on their plain embeddings), backend, and a terse human-readable summary. Retrieval itself is sub-second, so there is no retrieval progress to report.

Aggregate ingestion/enrichment status across all sources in a notebook — 'is my corpus ready'. Same rollup as source_status, plus source_count.

Hybrid search within a notebook: dense KNN + FTS5 lexical, fused with RRF. Returns raw chunks for the calling agent to synthesize. mode = hybrid | dense | lexical. Set private=true for a sensitive query so it is never dense-embedded via the remote backend (FTS5 is local and still runs); excluded remote spaces are flagged. Per-stage latency is returned under 'latency_ms'.