A local knowledge base plugin for DeepSeek Harness: It automatically organizes conversations into notes, methods, and pitfall cards saved as Markdown, uses local SQLite full-text search to allow the current AI to query via /wanxiang, and can also learn from folders bound to the conversation.
Stars 27
Reasons
- The content is readable and backup-ready Markdown, with an index that can be deleted and rebuilt; retrieval does not rely on embedding models, features tokenization and query enhancement optimized for Chinese, and reads content in three progressive layers (summary, overview, body) to control context usage.
- Supports importing Markdown, text, source code, folders, and ZIP files, with bidirectional links for Obsidian; includes a knowledge graph view.
- Automatic recall has three modes: off, prompt-only, and auto-inject. The author honestly notes that auto-injection is still experimental, with a miss rate of 76.25% on the synthetic strategy set.
Usage
- Download dsh-wanxiang-knowledge-0.2.0-alpha.10.tgz from Releases, install it via the local plugin package entry in Harness, or run dsh plugin --profile web add "<absolute path to plugin package>"; back up the knowledge directory before upgrading.
- Type /wanxiang <question> in a conversation to query existing knowledge; in a conversation bound to a folder, send /wanxiang study to learn from project materials.
Risks
- Currently in 0.2.0-alpha.10 preview, primarily validated only on Windows + Harness Desktop 0.2.1-alpha.1; does not parse PDF/Word body text, and lacks OCR and vector retrieval.
- AI organization and folder learning send relevant text to the model configured in Harness, incurring usage costs; optional knowledge content packs and security analysis extensions are geared toward reverse engineering and require separate installation.