dsh-kb-manager

Building a local RAG pipeline inside DeepSeek Harness usually means spinning up a vector database, an embedding service, and a retrieval middleware. dsh-kb-manager collapses that entire stack into a single plugin: multi-format import, CJK-aware smart chunking, SQLite-native vector plus full-text indexing, hybrid search, and citation tracing — all local, zero external processes.

The retrieval pipeline runs sqlite-vec KNN and FTS5 BM25 in parallel, merges via RRF, and optionally applies a reranker whose state is always returned explicitly as applied, fallback_rrf, or disabled, eliminating silent degradation. Imports and rebuilds are asynchronous with crash recovery and cooperative cancellation; deletes are millisecond soft-deletes that rebuild_index physically compacts, leaving no orphan vectors. All 22 Agent tools share a unified ok / data / error envelope, and every search result is flagged untrusted: true so the Agent always knows where data ends and instructions begin.

It suits developers and small teams working in DeepSeek Harness who deal with Chinese, Japanese, or Korean documents and want zero-dependency local retrieval. A single profile line gets it running offline, and the optional Web panel adds live import progress, hybrid search, and snapshot management with a visual interface.