dsh-knowledge

Project docs, personal notes, hard-won lessons from conversations, and bookmarked web tutorials tend to live in scattered places, leaving the model without context or verifiable citations when it answers. dsh-knowledge unifies all of that into a local SQLite FTS5 index, so the model retrieves relevant passages automatically and is forced to cite the source path, title, and line range for every hit. No cloud calls, no embedding APIs.

Three layers of capability work together. The document layer handles incremental imports of markdown, configs, and source code with Chinese trigram segmentation plus a short-term LIKE fallback for mixed scripts. The experience layer lets the model proactively distill real problems solved during a session into structured entries, and can backfill from historical session transcripts decoded via zstd. The web layer clips tutorials on demand, strips navigation and scripts, converts to markdown, and folds the result into the same search index. Outputting a kbquery fence renders an interactive search panel; a persistent star-map button in the bottom-right corner opens an Obsidian-style force-directed knowledge graph where nodes are color-coded by type, sized by content volume, and fully draggable with spring-linked neighbors.

This plugin suits developers and researchers who need a fully local, zero-cloud knowledge base and care about privacy, offline operation, and avoiding external API dependencies. Everything lives in a single file at ~/.dsh/knowledge/kb.sqlite for trivial backup, import paths are strictly confined to the workspace root, and binary or oversized files are skipped automatically, keeping the security boundary clean and predictable.