I gave a local LLM read access to my files, and it finds things better than Windows Search ever has

Many individuals rely on file titles and content for search, which can be problematic when providing vague keywords to Windows Search. While it attempts to match the exact term used six months ago, the results are often incomplete or filled with irrelevant web search outcomes. Local Large Language Models (LLMs) have proven adept at understanding text and determining its context, allowing them to effectively replace traditional search tools for many users.

By granting a LLM read access to specific folders and treating it as a search box, the experience has significantly improved for the author's local AI workflow. The official filesystem MCP (Model Context Protocol) server enables this functionality, allowing the model to list and read files within approved folders. Configured via a JSON block, the server interprets the specified command and arguments, granting access to the designated directory or drive.

For the author, it replaced Windows Search entirely, as the LLM demonstrated superior contextual understanding and more accurate search results. Although regular Windows search has improved with semantic indexing for Copilot+ PCs, the author's non-NPU-equipped i7-13700F F processor doesn't support it. Consequently, the author relies on regular search, which solely matches keywords, leading to issues with contextual search terms.

For instance, a note titled "night themes" would not appear in search results for "dark mode." Additionally, Start menu integration with Bing search further complicates the experience. The author's local filesystem MCP in LM Studio serves as a search alternative, comparing favorably to Windows Search. When prompted with "night theme," the LLM quickly identified the relevant file, while Windows Search returned unrelated results.

This pattern persisted across various search queries, highlighting the LLM's superior contextual understanding. However, limitations exist, such as the inability to open results directly within the LLM and potential speed issues on smaller GPUs. Despite these drawbacks, the author believes that Windows' semantic search will eventually improve and may rival the contextual search capabilities of local LLMs.