docindex

You know the feeling — dozens or hundreds of Markdown notes, PDF reports, and Office files piled up in a workspace, and the moment you need a specific conclusion you resort to Ctrl+F or manual browsing. docindex turns that into a single search box: it scans a directory you specify, builds a local semantic index over every document, and lets you ask in natural language or type keywords to get back matched snippets with exact line numbers and relevance scores.

Under the hood, retrieval runs two parallel tracks. Lexical search is powered by SQLite FTS5 with a built-in CJK n-gram tokenizer, so Chinese keyword search works without any native dependencies. Semantic search goes through a pluggable embedding slot — the default is a zero-dependency n-gram feature-hashing embedder that runs fully offline, and you can optionally switch to a Hugging Face transformers multilingual encoder for true cross-lingual recall. The two ranked lists are merged with Reciprocal Rank Fusion for a stable, interpretable order. Indexing is incremental by default with an optional file watcher; only changed files are re-processed, and a full rebuild is always available as a fallback.

It is built for two kinds of users. If you work inside DeepSeek Harness (dsh), the four model-facing tools — doc_scan, doc_query, doc_reindex, doc_stats — become available to the agent immediately, and any other plugin can call the same engine through the ctx.docIndex service. If you simply want local document search in your terminal, install it from npm and run the docindex CLI with no dsh environment at all. The core engine has zero runtime dependencies, making it a good fit for offline and privacy-sensitive setups.