Service 01
Enterprise retrieval over your own documents, policies and operational data, built so that an answer without a source cannot be produced. Most RAG deployments fail the same way: they answer confidently from general knowledge when retrieval returns nothing useful. The user cannot tell the difference. That failure mode is designed out of this one.
The Problem
The failure is not that the system is wrong sometimes. It is that a wrong answer looks exactly like a right one.
Failure 01
Retrieval returns three irrelevant chunks. The model answers anyway, from what it learned in training. The answer is plausible, fluent, and about a different company’s policy.
Our fix: retrieval confidence gates the answer. Below threshold, it refuses.
Failure 02
The answer is correct. Nobody can verify it, because nothing points back to the document it came from. Six months later the policy changes and nobody knows which answers are now stale.
Our fix: every claim carries its source document and section.
Failure 03
The system went live. Accuracy was never sampled. The first evidence of a problem is a customer acting on a wrong answer.
Our fix: monthly sampling and a report that includes the errors that reached a user.
The Build
The last two weeks are a supervised period where the system runs but is not yet trusted. That is not padding. It is where the accuracy baseline comes from.
Week 1
What documents exist, how current they are, what contradicts what. Most knowledge bases contain two versions of the same policy. We find those first, because retrieval cannot resolve a contradiction that the corpus contains.
Week 2
Structure-aware chunking rather than fixed windows. Section headings, tables and policy clauses survive as units. Freshness metadata attached at ingest, so staleness is queryable later.
Week 3–4
Hybrid retrieval, reranking, and the confidence gate that produces a refusal instead of an answer. The gate is tuned against your own corpus, because a threshold that works for legal documents does not work for support tickets.
Week 5–6
The system answers real questions, and every answer is reviewed before it reaches the asker. This produces the accuracy baseline, catches the corpus problems that only surface under real queries, and tunes the refusal rate.
Where It Fits
The pain is always the same shape: knowledge exists, but only in people’s heads or in documents nobody can find.
01
Tier-one support where the answer exists in a document but finding it takes longer than asking a colleague. Deflection is measurable from week one.
02
New joiners asking questions that are answered in a wiki nobody reads. The system does not replace the wiki. It makes the wiki answerable.
03
Exception handling, escalation rules and process edge cases that exist only in the memory of whoever has been there longest. The corpus audit usually surfaces that this knowledge was never written down at all, and that is itself the finding.
04
Where the evidence exists across four systems and assembling it takes a week. Source attribution is not a nice-to-have here; it is the deliverable.
Five to six weeks to production, with a supervised period before it is trusted.