mindspace-dsh-local-rag
适用于 DeepSeek Harness 的本地、由模型调用的混合 RAG
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Spirtxiaoqi7/mindspace-dsh-local-rag#1da456b164c7e60c0fc2622a7c8fbba31b256aaeMindspace Dsh Local Rag 以 mindspace-dsh-local-rag 发布，当前版本为 0.3.7。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
An independent, local-only retrieval plugin for DeepSeek Harness. It gives any tool-capable chat model one explicit search_local_memory tool instead of injecting retrieval into every turn.
This is the RAG branch of the Mindspace/DSH plugin family. Its retrieval design is based on lessons from ARPM, but it is not a direct transplant of ARPM's chat route:
defineTool mechanism so the active model decides when to search;compaction/end ingestion, per-session isolation, source provenance, document revisions, and a local model lifecycle;This repository is not coupled to mindspace-dsh-session-memory, and the two are not developed as one monolithic plugin.
| Capability | This repository: Local RAG | Structured-memory plugin |
|---|---|---|
| Responsibility | On-demand retrieval over large documents and older summaries | User profile, preferences, AI instructions, relationship mission, and roleplay preset |
| Model access | Standard tool calling | Governed personalization supplied to the active session |
| Package/repository | mindspace-dsh-local-rag | mindspace-dsh-session-memory |
| Persistence | <DSH_HOME>/mindspace-local-rag/ | Its own independent data store |
| Dependency | Does not depend on structured memory | Does not depend on Local RAG |
Local RAG can be installed or removed independently. On the DSH 0.1.1
compatibility line it composes with mindspace-dsh-session-memory: the two packages use
different Remote namespaces and do not share a service or data directory.
Conversation summaries in this RAG come directly from native DSH compaction
events; they are not read from the structured-memory plugin. Do not mix this
package pair with a legacy in-tree Mindspace memory checkout, because that
old integration owns a competing Memory service.
compaction/end summaries are deduplicated against older summaries and the recent uncompacted surface before indexing.local-rag://source/... address; the model can filter a second search by source/document or page through bounded extracted text without receiving a filesystem path.This README describes current main; it does not present an unpublished version
number as a Release asset. Requirements: Node.js 22.19+ (or 24+), pnpm, and a
local DeepSeek Harness checkout. Build a tarball from this repository, then
install it from the official Harness checkout root:
git clone https://github.com/Spirtxiaoqi7/mindspace-dsh-local-rag.git
Set-Location .\mindspace-dsh-local-rag
corepack pnpm install
corepack pnpm run build
corepack pnpm pack --pack-destination dist
$ragTgz = (Get-ChildItem .\dist\mindspace-dsh-local-rag-*.tgz | Sort-Object LastWriteTime -Descending | Select-Object -First 1).FullName
Set-Location C:\path\to\deepseek-harness
corepack pnpm dsh plugin --profile web add $ragTgz
corepack pnpm dsh --profile web --dump-config
corepack pnpm dsh web
Do not run pnpm dsh in the plugin directory: it belongs to the official Harness
checkout. Historic GitHub Releases map only to their respective tags and do not
represent the current main feature set.
Open Settings → Local RAG. Files can be uploaded and searched lexically before an embedding model is running. The first plugin install also obtains the Node ONNX native runtime; this is the model runtime, and it is not loaded during DSH cold start. The built-in verified model is shibing624/text2vec-base-chinese (ONNX, 768 dimensions, approximately 407 MB); the catalog is extensible but does not advertise unverified downloads.
The plugin stores its model and index beneath:
<DSH_HOME>/mindspace-local-rag/
Importing before the model is ready is safe. Downloading does not load ONNX or silently rebuild vectors; start the model explicitly and rebuild only when the settings page reports stale vectors.
The registered guidance is intentionally small:
Initial model retrieval needs only query and scope; a returned documentId or sourceId may be used for a same-source follow-up. Candidate counts, RRF parameters, output limits, filesystem paths, download sources, and model choice remain deployment-controlled.
Built-in Local RAG knowledge upload and retrieval entry
The settings page provides:
No remote embedding API is used. Network access is required only while downloading the pinned model artifacts.
pnpm install
pnpm run build
pnpm run test
pnpm pack --pack-destination dist
The suite covers real PDF/DOCX parsing, TSV provenance, deterministic RRF, scope isolation, lexical degradation, persistence/migration, committed compaction validation and deduplication, chunked upload safety, ModelScope-to-Hugging-Face fallback, resumable downloads, integrity and model lifecycle, strict Remote parity, and bounded tool output.
Current main has package version 0.3.7: it is qualified against DeepSeek
Harness 0.1.1-rc.2 and the stable 0.1.1 line in a clean profile alongside
DSH-memory 0.2.35. It
includes governed dual corpora,
startup reliability fixes, and the pnpm 11 no-install-script packaging repair. It
intentionally avoids reranking and user-configurable Top-K until retrieval quality
has been measured with real files, compaction summaries, and revision workflows.
MIT licensed.