Teaching AI to read our own documentation.
TracE’s hardest problem wasn’t the model — it was the knowledge. Thousands of IT articles were written for people, not retrieval. Processing and restructuring them at scale is what produced the house doctrine, and it was itself an exercise in AI leverage.
Reading the knowledge base at scale
Processed a body of IT documentation no manual review could keep pace with — surfacing what was outdated, duplicated, or unstructured, and restructuring 2,000+ articles into the retrieval-friendly, single-topic standard.
Curated a vetted POC corpus
While Confluence was being cleaned up, exported and vetted 100+ IT knowledge-base articles as a controlled corpus so the POC could prove accuracy on trusted content first.
Glossary & abbreviation generation
Authored glossaries and abbreviation guides across three Confluence spaces so retrieval resolves the acronyms and shorthand technicians actually type — a common failure point for naive RAG.
The doctrine that generalized
The hard lessons — structure beats prose, cite-or-abstain beats guessing — hardened into the house doctrine that anchors AutoBid’s Deviations Knowledge lookups and told the Scott AFB build never to infer a part number.