Finding the document, and what search cannot tell you

Two ways of finding a document, with opposite weaknesses

Keyword search matches strings. It finds "INV-2291" and every occurrence of "indemnity". It misses the document that says "holiday entitlement" when you searched for "annual leave".

Semantic search matches meaning, using the embeddings described earlier. It finds the holiday policy. It struggles with "INV-2291", because an invoice number is a string with almost no meaning and therefore no neighbours in meaning-space.

The two fail in opposite directions, and the new tools mostly do the second while the old tools do the first. Knowing which one you are using tells you which failure to expect.

The best systems do both and merge the results, usually called hybrid search. If yours does not, you have two searches to run.

No search tells you what it missed

A search that returns four results is not evidence that four exist. It is evidence that the system ranked four highly enough to show you. This is obvious when stated and forgotten constantly, especially when the answer arrives in a fluent sentence rather than a list of links.

For "is there anything about X?", one search is never enough:

• Search in your words.
• Search in the organisation's words — the actual phrase your policies use.
• Search for the person who would have written it, or the year.
• Search filenames as well as contents.

If four different approaches turn up nothing, you have weak evidence of absence. One search turning up nothing is worth nothing.

Relevance is not authority

Ranking answers the question "which document best matches this query". It does not answer "which document is currently in force".

So a superseded 2019 policy whose wording happens to match your phrasing more closely will rank above the 2026 replacement that uses different words. It is complete, plausible, professionally formatted and wrong, and every visible signal — the tone, the letterhead, the structure — is identical to the right one.

The defence is a habit, not a setting: check the date and the version of anything you are going to rely on, before you read it properly. Look at the file's date, then look inside for a version or review date, then check whether a newer one exists in the same location.

What actually determines whether search works

Uncomfortable, and true: search quality in most organisations is set by the documents, not the tool.

If your estate is full of Final v3 (2) FINAL.docx, documents with no titles, no dates in the text, no owner and five near-identical copies in different folders, no model repairs that. Semantic search will confidently retrieve one of the five copies, and there is no principled reason it will be the right one.

The cheap improvements, in order of value: put the date and status in the first line of the document itself, keep one authoritative copy and delete or clearly mark the rest, and use file names a stranger could understand. All three help humans as much as machines, which is how you know they are real improvements rather than preparation for a tool.

Ask for the source, always

Where a search tool answers in prose rather than links, the same demand as everywhere else:

Answer only from the documents. After each statement, give the file name, the date of the document, and the sentence you took it from.

Then check that the date is recent enough and the sentence really appears. Two seconds each, and it converts search output into something you can put in front of somebody.

The free path, and it is excellent

For finding things on your own machine or a shared drive, free tools comfortably beat most paid ones:

• ripgrep (rg) searches the full text of an enormous folder tree in seconds, from the command line, with exact matching and regular expressions. It is the fastest way to answer "does this exact phrase appear anywhere".
• Recoll is free and open-source, indexes documents, PDFs, spreadsheets and mail on Windows, macOS and Linux, and gives you full-text search with a normal interface.
• DocFetcher does the same, cross-platform and portable.
• Everything on Windows searches file names instantly across every drive.

Add a semantic layer only if you need it, and it is also free: a local embedding model through Ollama over your own folder. Start with rg and Recoll, though. Most of what people want from "AI search" is full-text search over documents that were never indexed at all.