AlowEnsoul/dsh-rag-kb--packages-kb-rag-kb ↗★ 2
@deepseek-ai/dsh-rag-kb
本地RAG知识库宿主插件：Ollama嵌入、JSON持久化与检索工具。 适合需本地知识库检索的用户；依赖本地Ollama嵌入服务。
安装
此插件尚未提供可验证的 bundle，或兼容性检查未通过。请先阅读仓库说明。 阅读完整 README ↗
说明文档
阅读完整 README ↗@deepseek-ai/dsh-rag-kb
Local RAG knowledge base host plugin.
- Persists named knowledge bases to
dsh-kb-store.jsonin the workspace root. - Computes chunk embeddings through a local Ollama
/api/embedendpoint (configurable model/URL/chunk size). - Serves a single
/kbJSON RPC route ({ method, ...args }) for the browser UI. - Registers three model-facing tools:
kb_search,kb_index,kb_status.
Services
- Consumes
webServer,fs,subprocess,sandboxPolicy,tools.
Model Experience
Request context and condition
What the model sees
Three tools: kb_search, kb_index, kb_status (see the generated tool catalog).
Token effect
Conditional: tool results appear only when the model calls these tools.
KV Cache effect
Does not invalidate the reusable prompt prefix; tool results are request-local.
Known Limitations and Deferred Work
- Embedding batches use a single local Ollama endpoint; no queueing, retry, or concurrency limits beyond a per-request
--max-time. - Changing the embedding model after indexing mixes vector spaces in one knowledge base.