An AI-powered motorcycle service manual search tool. Parses chunked service manual data and expands part-name aliases to instantly surface torque specs, procedures, and references from motorcycle pdf service manuals.

A service manual is the original RAG problem. Unfortunately, the industry-standard format, the PDF, is actively hostile to retrieval. Searching through 500+ page manuals is painful, and users have to remember manufacturer-specific synonyms for part names. The information they need most, such as torque values, clearances, and procedures, is buried inside un-linked tables, diagrams, and layouts that off-the-shelf parsers often destroy. The hard part is not just the retrieval layer. It is the authoring pipeline: turning a cursed, print-centric format into structured, typed, citation-grounded chunks that a vector store can serve quickly and reliably.

Off-pipeline, low-confidence, and ambiguous extractions are routed into a review queue prioritized by a composite risk score combining retrieval confidence and keyword-match rank. Usage patterns and reviewer decisions, including validations, rejections, edits, parse corrections, and synonym updates, are persisted as structured, versioned ground-truth data for reproducible A/B evaluation across models, embeddings, and extractors, while building a labeled corpus for downstream fine-tuning.

To solve the trust problem in AI systems, every answer is source-linked. This is inspired by my work building AI tools for cancer diagnosis: turning long, structured, technical documents written for experts into fast, navigable, accountable references for clinicians who need answers in context. The patterns transfer: provenance, confidence, fallback, and structure. The domain is different. The problem shape is the same.