Use case · Knowledge retrieval
A knowledge layer your teams can actually trust.
A token factory for retrieval over the corpus the organisation already owns — wikis, runbooks, product specs, incident post-mortems, policy libraries — without handing the knowledge to whichever vendor ran the onboarding demo best.
The problem
What we are actually solving.
Every mid-market organisation has a knowledge problem, and almost none of them are about knowledge. They are about trust. The wiki is stale in patches. The runbooks live in three formats. The definitive answer to “how do we handle this” lives in a Slack thread from two years ago. A search box helps the people who already know where to look; it does not help the people joining next Tuesday.
The tempting buy here is a vendor retrieval product that ingests everything, indexes it, and answers questions. The trap is that the ingested corpus is often the most valuable asset the company has — the collected judgement of ten or twenty years — and the vendor contract usually does not reflect that. You end up renting the ability to search your own knowledge, on an index you cannot move.
The approach
The shape of the token factory for this use case.
The context engine is the full corpus, ingested on a schedule that matches how fast the source changes. Wiki-style content updates continuously; policy documents update rarely and formally. The ingestion pipeline handles both without pretending they are the same.
Chunking respects source structure — headings, callouts, code blocks, policy sections. Every chunk carries provenance: which document, which revision, which owner, last reviewed when. The last-reviewed metadata becomes part of the retrieval signal: answers sourced from a stale runbook surface a warning, not a confident paragraph.
The prompt compiler is deliberately small here, because the intents are narrower than people expect. Most knowledge retrieval collapses to four shapes: “what is our position on X,” “where is the procedure for Y,” “who owns Z,” “has this been seen before.” Four intents, well-compiled, beat a hundred intents done casually.
The router favours models with strong instruction following over raw reasoning budget; knowledge retrieval failures are rarely about reasoning, they are about restraint — refusing to synthesise when the evidence is thin. The eval harness tests refusal as a positive outcome, not only correct answers.
Outcomes
What changes.
Built well, an internal retrieval factory shifts the organisation from “ask someone senior” to “ask the factory, then escalate the genuinely novel question.” Onboarding gets measurably shorter; tribal knowledge gets surfaced instead of lost; the people who know things stop being bottlenecks.
The trade-off: a retrieval factory is only as good as the editorial discipline around the corpus. If the wiki is half-abandoned when the factory is built, the factory will surface the abandonment. That is often the most useful thing it does in the first quarter — but it is not what the business case usually promises.
Light regulatory frame. GDPR applies to the corpus as soon as it contains personal data, and most internal corpora contain more than their owners think. The factory’s own logs are themselves a processing activity. Build with data residency and retention as first-class parameters, not afterthoughts.
Honesty
Expert approach based on adjacent delivered work. I have not shipped a production knowledge-retrieval token factory under Blue Abraham; the approach above is how I would build it given what I delivered in aviation technical documentation. See Technical documentation for the delivered parallel.
Aviation technical documentation is, structurally, a very hard knowledge-retrieval problem: long, formal, cross-referenced, with a regulatory bar on answers. The retrieval patterns that worked there transfer cleanly to enterprise corpora — the gap is editorial, not architectural.
Next step
Talk about a knowledge layer your teams can actually trust..
The usual first engagement is an architecture review. Say so on the form, or write directly.
Get in touch