Where do we get the most benefit for the least work?

Filtering by recency and by owning team, if those fields are trustworthy. Most poor answers in an internal corpus come from superseded material rather than irrelevant material, and both look equally good to a ranking system, so a filter applied before ranking removes a whole category of failure at almost no cost.

Our fields are inconsistent. Is it still worth it?

Yes, but scoped to where they are reliable. A department that maintains its fields properly can have filtering switched on for its material while the rest of the corpus does not, which is far more useful than waiting for an organisation-wide tidy-up that will not happen. Partial coverage still removes real failures.

Why does a document I wrote not appear for a colleague?

Because access is recorded here too, and applied per person at the moment of asking. If the fields recorded when it was taken in say the material is restricted, that is what governs, regardless of what the source system says today, until the next refresh brings the two into step.

Glossary

Metadata

Metadata is the set of facts about a document rather than in it, and filtering on those facts removes more wrong answers than improvements to ranking usually do.

In plain terms

The facts about a document rather than the words inside it: who owns it, when it was last touched, which team it belongs to, whether it was ever approved. Using those to narrow a search removes a great deal of wrong material very cheaply, which is why it is the most under-used tool in this whole area.

Why it matters

Because the most common bad answer is not irrelevant but out of date, and no amount of ranking work distinguishes a current policy from the version it replaced. Both are equally relevant to the question and equally well written; the only thing separating them is a fact stored beside them, and filtering on that fact costs almost nothing.

How it works

It travels alongside the material rather than inside it, which is what makes it usable as a filter. Because owner, date, department and status live as separate fields, a search can restrict itself before ranking anything, so the pool being ranked is already free of the superseded and the irrelevant rather than being sorted with them still in it.

The date field is the single highest-yield one and it is frequently wrong. Many systems record when a file was last touched rather than when its content was last reviewed, so a document reformatted last week looks fresher than one carefully revised two years ago, and filtering on a date that means the wrong thing produces confident staleness.

Access rules are metadata too, which is why this apparently administrative subject turns out to be load-bearing. What each person may see is recorded beside the material and applied per person, so the same fields that let you filter for recency are the ones deciding whether somebody is shown a document at all.

Its quality is set by processes nobody designed for search. Fields get filled in by whoever saved the file, under whatever conventions their team happened to adopt, so a corpus can be perfectly organised in one department and blank in another, and search behaves accordingly without ever explaining why.

Two documents a ranking system cannot separate

This is the argument for a subject that sounds like filing. When people describe internal search as unreliable, the examples they give are very often of this shape: the system found something genuinely relevant that happened to be obsolete, and presented it with the same confidence as anything else. It is worth being precise about why better ranking cannot help. Ranking orders documents by how well they answer the question, and the superseded version answers it beautifully, because it was written to. The information that separates them is not in the text at all; it lives in a field recorded beside the document, and a filter applied before ranking removes the whole category in one move. That is why this is the cheapest lever available and why it is so consistently skipped, since it looks like administration rather than like improving search, and the people who would benefit are not the people who maintain the fields.

Seen in the wild

A curated knowledge layer that flags stale material and identifies gaps, which is a judgement made on the facts about a document rather than its words.
Guru
Enterprise search that shows each person only what their permissions allow, which is a filter applied on the facts recorded beside each document.
Glean
A workspace where pages and databases carry properties alongside their content, so the facts about a page are structured by default.
Notion AI

Common misconceptions

People assume

It is administrative housekeeping.

In fact

It is the cheapest available improvement to answer quality and it also enforces who sees what. Filtering out the superseded before anything is ranked removes a class of wrong answer that no amount of work on ranking or on the model addresses, because both candidates are genuinely relevant.

People assume

The dates we have are the dates we want.

In fact

Most systems record when a file was last modified, which changes when somebody fixes a typo or converts a format. That is a poor proxy for whether the content is still correct, so filtering on it can surface a cosmetically recent document over a substantively current one.

Telling them apart

Metadata vs Knowledge graph

Metadata

Facts about each document, held beside it.

Knowledge graph

How things relate to each other, held as connections.

One filters a set of documents and the other answers questions that span several, which is why a graph is much more work and answers a different kind of question.

Questions

Where do we get the most benefit for the least work?: Filtering by recency and by owning team, if those fields are trustworthy. Most poor answers in an internal corpus come from superseded material rather than irrelevant material, and both look equally good to a ranking system, so a filter applied before ranking removes a whole category of failure at almost no cost.
Our fields are inconsistent. Is it still worth it?: Yes, but scoped to where they are reliable. A department that maintains its fields properly can have filtering switched on for its material while the rest of the corpus does not, which is far more useful than waiting for an organisation-wide tidy-up that will not happen. Partial coverage still removes real failures.
Why does a document I wrote not appear for a colleague?: Because access is recorded here too, and applied per person at the moment of asking. If the fields recorded when it was taken in say the material is restricted, that is what governs, regardless of what the source system says today, until the next refresh brings the two into step.

Key takeaways

Filtering before ranking removes the superseded, which ranking cannot.
The date field usually records modification, not review, and that matters.
Access rules are metadata, which is why this is load-bearing rather than clerical.
Quality is set by habits that were never designed with search in mind.