AI Search

Semantic search over your own documents and data.

Nothing launched in this category yet.

Be the first

About AI Search

Search over your own material is the most quietly useful application of language models, and it is also the one where the difference between a good and a bad implementation is most visible. The promise is simple: ask a question in your own words and get an answer from your documents, your notes, your codebase or your company's accumulated writing. The failure mode is equally simple — a confident answer that is not in any of those places.

The single most important feature in this subcategory is citation. A tool that shows you which document and which paragraph an answer came from is one you can use for work that matters, because you can verify in five seconds. A tool that returns a fluent paragraph with no provenance is asking you to trust it on questions you asked precisely because you did not know the answer. Given two otherwise similar tools, take the one that cites.

The second thing that separates implementations is what happens when the answer is not there. Good systems say so. Weak ones assemble something plausible from adjacent material, which is worse than a null result because it is harder to detect. Test this deliberately: ask something you know is absent and see what comes back.

Beyond that, the practical differences are about ingestion and freshness. What formats does it read — PDFs with tables, scanned images, spreadsheets, code, email? How does it handle a document that changed yesterday? Systems that reindex continuously stay correct; ones that index on upload quietly drift out of date, and the answers get worse in a way nobody notices for weeks.

Where the index lives is the privacy question. Some tools embed and store your documents on their servers, some keep the index local and only send the query, and some run the whole pipeline on your machine. All three are reasonable choices for different material; what is not reasonable is not knowing which one you picked. For local options see [local models](/categories/local-models).

If you are evaluating one of these for a team rather than yourself, permissions are the thing that will bite you. A search system that ingests everything and answers everyone will happily surface the salary spreadsheet to whoever asks the right question. Tools that respect the source system's access controls at query time are doing meaningfully harder engineering, and it is worth paying for.

From the blog

Reading on launching, ranking and ai search.

All posts