The parser solves the data problem head-on: Textract classifies every region of every page, tables become their own CSV files, figures are pulled by exact bounding box and read by a vision model (diagrams → Mermaid), and each document lands as one clean, normalized folder. Everything else stands on it.

Tables never get inlined, since that burns context and drowns the search. Each stays a file behind an inline tag the chunker can't split, and an MCP server turns it into a tool the model queries at runtime to filter, aggregate, and join real rows.

Every question is routed by cost: a lookup answers from a template in ~200 ms; a real one runs the full path. Three-layer discovery (keyword catalog → summaries → filtered hybrid search), RRF fusion, optional Cohere rerank, compose, then verify the answer against its own evidence. Feedback tunes the ranking.

The recent push was scale: from a 5-document prototype to a real company SharePoint crawl. One orchestrator now handles local files and SharePoint, skips unchanged content, keeps moved documents attached to the same identity, and streams crawl progress into the console.

I run it like production, not a demo: every sub-project goes spec → adversarial review by a second model → plan → TDD, one commit each, and that reviewer catches a real issue at every gate. The docs say plainly what's proven by tests versus what hasn't had a real end-to-end run.

The console around it: a Svelte streaming chat (live pipeline · evidence · citations), the live SharePoint corpus tree, an upload → parse → index pipeline, and a per-query inspector. Plus RAGAS + side-by-side eval, LangSmith tracing, and every reported bug locked behind a regression test.