Releases
v1.7.0
context-mode v1.7.0 — Local Code Indexing
Compare
Sorry, something went wrong.
No results found
What's New
Local Code Indexing (Merkle diff + hybrid search)
Merkle-style incremental indexing — SHA-256 + mtime diff. Re-indexes only changed files.
Semantic chunking — tree-sitter AST extraction (functions, classes, interfaces, methods) with tiktoken fallback for unsupported languages.
FTS5 + vector hybrid search — BM25 full-text ranking fused with cosine similarity from local embeddings.
ONNX / Ollama / zero-vector fallback — all-MiniLM-L6-v2 via onnxruntime-node, Ollama API, or graceful FTS5-only degradation.
Cross-encoder reranking — optional ms-marco-MiniLM-L-6-v2 ONNX reranker + heuristic scoring fallback.
File watcher — fs.watch with 1-second debounce; persists watch list to ~/.context-mode/watched.json.
CLI commands : local-index, local-search, local-repos, local-status.
MCP tools : ctx_local_index, ctx_local_search, ctx_local_status, ctx_local_repos, ctx_local_watch, ctx_local_unwatch.
WAL-safe SQLite schema — dedicated code-index.db with FTS5 virtual table, vectors BLOB store, and jobs tracking.
Technical Details
384-d embeddings, batch size 32.
Tiktoken truncation for result formatting.
Lazy model download via .url markers in ~/.context-mode/models/.
WAL mode + mmap_size = 256MB for read-heavy search workloads.
You can’t perform that action at this time.