ahlanlhw/cos

COS, short for Cognitive Operating System, is a local, offline search tool designed to help AI agents quickly find relevant information inside large markdown knowledge bases such as wikis, documentation sites, and runbooks. The core problem it solves is that when an AI agent needs to orient itself in a large knowledge base, reading every file from scratch is slow and consumes a lot of the agent's limited working memory. COS pre-indexes all the markdown files and makes them instantly searchable. The tool is built on SQLite, a lightweight file based database that runs entirely on your machine with no internet connection, no API keys, and no external services required. It uses FTS5, a built-in SQLite feature for full text search, to scan across titles, page bodies, file paths, and tags. When ranking results, it blends several signals: BM25, a standard keyword relevance algorithm, a wikilink graph score that treats pages with more incoming links as more authoritative, a recency boost for recently updated pages, and a confidence score from frontmatter metadata. Every component of the final score is returned in the output so you can see exactly why a result ranked where it did. A key feature is context packs, a command that assembles the top-N most relevant pages with excerpts into a single file ready to hand to an AI agent as a prompt attachment. There is also a duplicate detection feature that flags near-duplicate pages for review. You would use COS when building or running AI agents that work with a growing markdown knowledge base and need fast, explainable retrieval rather than raw file reads. It is written in Python with a command-line interface.