02 / Retrieve / CL650 manual retrieval
Technical answers, with the sources attached.
A local retrieval system that helps Claude answer questions from Challenger 650 simulator manuals without relying on memory.
Working retrieval system · flight simulator study only, not real-world operations
- manuals indexed
- 21
- retrievable chunks
- 11,362
- noisy chunks removed
- ~4,000
01 / The brief
Start with the source.
The collection spans operating manuals, quick-reference handbooks, avionics guides, limitations, weight and balance, and training material. Claude is instructed to treat those documents as the authority: retrieve first, answer from the returned passages, cite the manual and section, and say when the evidence is insufficient.
02 / The system
A local pipeline from PDF pages to an MCP retrieval tool.
OpenDataLoader parses the PDFs inside a temporary Docker container. A structure-aware chunker preserves tables and page context. Document chunks and queries go to Voyage for embeddings; LanceDB stores the resulting search index locally. An MCP server gives Claude a focused search tool, returning relevant passages from the library when needed.
03 / The hard part
Document cleanup mattered as much as model choice.
The source material contains jumpy heading levels, repeated watermarks, page furniture, dense numeric tables, and cockpit diagrams that parse into word salad. Targeted filters removed roughly four thousand low-value chunks while protecting real tables. A three-page scanned limitations document needed a separate vision transcription pass with explicit instructions not to invent unreadable values.
04 / What remains
The most valuable charts are also the clearest boundary of the current system.
Some performance data exists only in tiled raster charts, so today the index can retrieve captions without reliably retrieving the plotted values. The proposed next phase groups image tiles by page, renders the original page, classifies tables and graphs, uses tiered vision extraction, labels graph readings as approximate, and requires a human-reviewed manifest before processing the corpus.