mindspace-dsh-local-rag

Traditional retrieval-augmented generation (RAG) methods often force retrieval results into every chat turn, which can lead to redundant context or irrelevant information interference. The mindspace-dsh-local-rag plugin solves this by providing an explicit search_local_memory tool, enabling models to autonomously decide when to search the local knowledge base only when the conversation context is insufficient, resulting in more precise and controlled retrieval.

The plugin's core capabilities lie in its local dual-corpus management: it fuses vector similarity and BM25+ lexical search using RRF algorithms, manages two corpora—user-uploaded documents and session compaction summaries—and handles model lifecycle entirely locally without external APIs. It ensures secure, traceable retrieval with support for document parsing, revisions, and provenance tracking.

mindspace-dsh-local-rag is ideal for DeepSeek Harness users, especially developers or researchers who prioritize data privacy and prefer local RAG deployment. It operates independently from structured memory plugins, allowing for standalone installation or removal, making it a flexible solution for those seeking autonomous and customizable retrieval capabilities.