AWS says fix document extraction before you touch retrieval

AWS published a Textract-Bedrock walkthrough that runs documents through Textract before they reach a Bedrock knowledge base, using utility bills as the worked example. The argument is that preprocessing, not a better embedding model, fixes retrieval on complex documents.

Utility bills are tables and totals arranged spatially, and a naive PDF-to-text pass flattens that layout into a wall of unstructured text. Once the structure is gone, retrieval tuning can't bring it back, because the model can find the right chunk and still answer wrong.

I hit this building Rayni. Retrieval usually worked, the right chunk came back, and the extraction was wrong anyway. A table had already been mangled into prose before it reached the vector store, and no amount of reranking brought the structure back.

Textract's job is layout: recognizing totals and line items instead of just pulling characters off a page. That's a computer vision problem, and RAG pipelines that skip it are asking one model to do work built for two.