Recall
Recall ranks captured sources by semantic similarity to your query:
khiipd recall "Inca knot record system"khiipd recall "building a second brain with AI" --limit 5How it works
At v0.1.x, recall uses the bundled MiniLM-L6 ONNX model to embed a per-source embed-text composition of each capture’s typed payload — a deterministic projection of the structured fields that matter for that source (title, body, author, key entities) rather than a raw text dump. Queries are embedded the same way, and recall returns cosine top-k matches (per ADR-0009 §C7).
This runs locally, at zero LLM cost, and works offline after the one-time model fetch on first use. Quality is whatever MiniLM-L6 gives you — good enough for “find that thing I captured about X,” and pluggable for more (below).
Tuning
--limit N— how many results to return (default 10).- Recall is by meaning, not keyword — “how LLMs actually work” can surface a captured talk transcript and an article that never use those exact words.
Pluggable embedders (roadmap)
The embedder sits behind a Protocol so it can be swapped without touching the rest of the substrate. Planned for v0.5+:
- Local LLM via Ollama (better quality, still free + local)
- BYOK (OpenAI / Anthropic / Gemini) for best-quality embeddings
- BM25 keyword fallback for exact-term recall
Switching embedders requires a re-embed of the corpus (a backfill); the captured payloads themselves don’t change.