dsh-llm-wiki

In agentic workflows, the main model frequently faces a sprawl of Markdown and plain-text files with no unified, searchable, and source-attributable knowledge base. dsh-llm-wiki closes that gap by isolating data under the canonical realpath of the active workspace, importing workspace documents into local storage, building a disposable PGlite index, and handing the agent a full toolchain for search, read, write, and graph exploration, all without any hidden nested model calls so that every model input and output stays in the normal session log.

At the core, the plugin imports .md, .markdown, and .txt files into raw/docs and constructs a hybrid retrieval pipeline that fuses 384-dimensional feature-hashing embeddings, PostgreSQL full-text search, a CJK substring fallback, and [[wikilink]] graph signals through reciprocal-rank fusion. Search results carry plugin-relative paths, source line ranges, excerpts, the fused score, and each signal's contribution. The agent can create or update sourced Wiki pages under wiki/sources, wiki/entities, or wiki/concepts via llm_wiki_write, with every replacement guarded by a read-before-write SHA-256 digest to prevent accidental overwrites. llm_wiki_graph lets the agent explore entities, mentioning pages, relations, and paths of up to three hops. The index directory is safe to delete at any time and is fully rebuilt on the next llm_wiki_ingest, while the Markdown files remain the single source of truth.

The plugin is built for DeepSeek Harness developers who manage project documentation, research notes, or concept wikis in local Markdown and want the agent to cite precise line ranges, read from and write back to sourced pages, and keep all data strictly within the local workspace. This release is intentionally an MVP: it supports local UTF-8 text only, ships without a Web UI, MCP server, PDF parsing, or background model tasks, and prioritizes predictable offline behavior and per-workspace data isolation.