The Outline Shape

AutoFix has a library of documents - service manuals, technical bulletins, and recall notices - and your agent needs to find its way around them.

The reflex move is to chunk every document, embed the chunks, and search by similarity. That hands the agent a slice that reads like the query. But a slice has no address: the agent can't tell which document it came from, which section it sits in, or what comes before and after it. And the cut itself is lossy - a fixed-size chunk boundary can fall mid-procedure, and once a fact is split across two chunks at ingest, no reranker or bigger context window puts it back together. (see Fixed-size chunking destroys structure before retrieval ever runs - see PageIndex on the "chunking floor.")