How mertvbaran/recall works
Recall is a local, retrieval-only search tool for a personal Markdown wiki: it returns ranked passages and sources, and can flag questions outside the notes rather than generating answers. Indexing parses vault documents, chunks them, embeds them, and persists a searchable index; querying embeds the question, retrieves candidates, and applies a separate confidence gate. Users interact through a CLI or Streamlit dashboard. The dashboard also records gate corrections and presents evaluation reports. Pipeline and interface relationships are documented; some implementation details are unsampled, so uncertain wiring is omitted.
Main entry points: build_index.py Index builder · query.py Query CLI · dashboard.py Streamlit dashboard · embedder.py Embedder
Components 15
Indexing
- Index builder
src/build_index.py - Vault loader
src/loader.py - Chunk strategies
src/chunkers.py - Embedder
src/embedder.py
Retrieval
- Vector search
src/store.py - Scope confidence gate
src/reranker.py
User interfaces
- Query CLI
src/query.py - Streamlit dashboard
src/dashboard.py - Gate feedback log
src/feedback.py
Local data and models
- Foundry Local
- Embedding cache
src/embedder.py - Vector index
src/store.py - Evaluation reports
results
Other components
- Markdown vault
- User
Connections 17
- Markdown vault → Index builder: indexes
- Index builder → Vault loader: loads documents
- Index builder → Chunk strategies: chunks documents
- Index builder → Embedder: embeds chunks
- Index builder → Vector index: persists index
- Embedder → Foundry Local: requests vectors
- Embedder → Embedding cache: reads/writes vectors
- User → Query CLI: asks question
- Query CLI → Vector index: loads index
- Query CLI → Embedder: embeds question
- Query CLI → Vector search: searches vectors
- Query CLI → Scope confidence gate: checks candidate scope
- User → Streamlit dashboard: asks question
- Streamlit dashboard → Vector search: searches vectors
- Streamlit dashboard → Scope confidence gate: checks candidate scope
- Streamlit dashboard → Gate feedback log: records corrections
- Streamlit dashboard → Evaluation reports: reads reports