Amazon Bedrock Knowledge Bases is AWS's managed retrieval-augmented generation (RAG) service: it ingests your documents, indexes them, searches them at query time, and generates cited answers. Most attention goes to the generation side — and Claude is explicitly supported there; AWS's own RetrieveAndGenerate examples pass Claude model ARNs. But answer quality is capped by ingestion quality: if the pipeline mangled a table or skipped a diagram caption during indexing, no generation model can recover the lost information. That is where the parsing configuration comes in.
The default path: text extraction
By default, a knowledge base ingests documents by extracting their text, then chunking it for embedding and indexing (default chunking produces roughly 300-token chunks that honor sentence boundaries; fixed-size, hierarchical, semantic, and no-chunking strategies are also available). Plain text extraction is fast and cheap, and for prose-heavy documents it is usually enough. Its blind spot is everything that isn't linear text: multi-column layouts read out of order, tables flattened into meaningless word streams, and figures, charts, and scanned pages contributing little or nothing.
The advanced path: a foundation model as parser
Knowledge Bases also supports advanced document parsing, where a vision-capable foundation model reads each document and produces the text representation that gets chunked and indexed. AWS's supported-models documentation explicitly lists Claude vision models among the model families usable as a parser, alongside Amazon Nova vision models and Llama 4 vision models. The same page notes that inference profiles are supported for parsing (and response generation), enabling cross-Region inference — with AWS's caution that cross-Region inference can move data between Regions, which your data-residency review should note.
Why does a vision model make a better parser? Claude's document processing analyzes page images and text together — the same capability that powers its PDF understanding on the Anthropic API, where every PDF page is processed both as an image and as extracted text. Applied at ingestion time, that means:
- Complex layouts — multi-column pages, sidebars, headers and footers — can be read in logical order rather than raw extraction order.
- Tables can be rendered into text that preserves row/column relationships instead of a flattened token soup.
- Mixed-content documents — reports where charts, figure captions, and prose interleave — yield a text representation that describes visual content, so the information becomes searchable at all.
The trade-off is straightforwardly economic: parsing every page through a foundation model costs foundation-model invocations, and ingestion of a large corpus becomes a meaningful line item where default extraction was nearly free. It also runs slower. Reserve it for collections where layout actually carries meaning.
Configuring it
Parsing is part of the knowledge base's data-source ingestion configuration, set when you create or update the data source — you select advanced parsing and specify the parsing foundation model (a Claude vision model or inference profile). The exact console fields and API parameters evolve; check the AWS Knowledge Bases documentation for the current configuration shape rather than hard-coding assumptions. Note two adjacent facts: chunking strategy is configured separately and applies to whatever text the parser emits, and for audio/video content AWS uses the Bedrock Data Automation parser to convert media to transcripts and scene summaries before standard text chunking.
How this fits the bigger Bedrock picture
Knowledge Bases runs on Bedrock's own runtime APIs with anthropic.-prefixed and ARN-style model identifiers — not the Anthropic Messages API surface. Anthropic's server-side web search and web fetch tools are unavailable on Bedrock, so Knowledge Bases is the Bedrock-native grounding path, and the parser choice is one of its main quality levers. A fully Claude-flavored pipeline on AWS can use Claude vision for parsing and Claude for generation, with AWS-side embedding models (Amazon Titan or Cohere — Anthropic doesn't offer embeddings, and Voyage models are not documented as KB embedding options) in between.
Where to go next
See chunking strategies for what happens to parsed text next, Knowledge Base embeddings for the indexing layer, and RetrieveAndGenerate for the query side.