What happened
Cohere published a guide describing how to build a financial-research assistant on AWS for questions over financial filings. The stated use case is familiar to analysts: locate a number in a long filing, determine whether it came from a table, chart, footnote, or prose, and show enough evidence that another person can check it. Cohere frames this as harder than ordinary document search because PDF extraction can scramble multi-column layouts, remove charts, and destroy the row-and-column relationships that give a table its meaning.
The proposed system assigns distinct jobs to four model components. Cohere Embed converts each page and each question into vectors for semantic retrieval. The guide uses Embed Pro for offline document indexing and Embed Fast on the live query path. After vector search finds candidate pages, Cohere Rerank can reorder them; the example selects Rerank Pro because retrieval accuracy is prioritized over response speed for this use case. Its Context Focus feature identifies passages judged relevant to the question. Cohere Parse is used for pages with weak text extraction or table-heavy layouts, returning Markdown with tables represented as rows and columns.
Anthropic Claude on Amazon Bedrock generates the final answer from the selected evidence. The web application has three described modes: an assistant that answers questions with citations; a search view that can compare results with and without reranking and display Context Focus passages; and a Parse view for inspecting extracted tables and converting scanned-looking pages into text. A citation identifies the source page and the type of evidence on it, and the interface can open the page image for verification.
The architecture separates ingestion from answering. During ingestion, Parse runs first only where needed, then Embed Pro indexes page text and stores vectors in Amazon OpenSearch Serverless. At question time, Embed Fast embeds the query, nearest pages are retrieved, Rerank may reorder them, and Claude writes from the resulting material. Cohere models run as real-time endpoints on Amazon SageMaker AI, deployed through AWS Marketplace; the application, sign-in, API, and ingestion pipeline use serverless AWS services.
The guide also makes explicit cost and latency choices. It recommends 1,024-dimension int8 vectors, optional Parse and reranking stages, and applying Parse only once to pages where it can improve extraction. For long pages, it sends Claude selected passages together with the page image: the passages reduce input volume while the image preserves layout context. It also proposes multilingual embeddings rather than a separate translation step for questions and documents across languages. A CloudFormation-based demo is available for deployment in an AWS account, but the source does not state a price for the complete implementation.
What the archive adds
This is a new implementation example on a related enterprise-data thread. Earlier Field Notes on AtScale and Amazon Bedrock AgentCore and Strategy Mosaic and Amazon Bedrock AgentCore examined governed semantic definitions as a boundary between agents and enterprise analytics. Cohere’s example addresses a different layer: recovering and exposing page-level evidence from source documents.
Trend signal
The design supports verification bandwidth is the scarce engineering resource. The system is not presented as a way to eliminate analyst review. Instead, it makes review cheaper by showing the exact page, preserving page imagery, labeling the evidence type, and reducing the material a reviewer must inspect. Reranking and parsing matter here because they improve the route to verifiable evidence, not merely answer polish.
Why it matters
Teams often treat retrieval quality as an internal technical metric and citations as a presentation feature. This guide treats both as part of the product contract. In finance and similarly regulated or high-consequence domains, an answer without an inspectable source may be unusable even if it is correct. The useful system boundary is therefore not “an LLM connected to PDFs,” but an evidence path from a user question to a specific source page.
The implementation also gives teams a way to separate quality investments by frequency. Expensive indexing quality can be paid once; query-time retrieval can be optimized for latency; parsing and reranking can be reserved for cases where they change the result. That does not prove these particular services are optimal for every corpus, but it is a practical decomposition for evaluating any document assistant.
Limits and uncertainty
This is a Cohere-authored build guide, not an independent evaluation of retrieval accuracy, citation correctness, latency, or total operating cost. It does not report benchmark results against alternative parsers, embedding models, rerankers, or answer models. The source also does not establish how well the design handles all filing formats, whether citations remain correct under adversarial or ambiguous questions, or what human review policy is required before use in investment workflows. The CloudFormation demo shows deployability, not production suitability.
Skill to build now
Take a representative set of your own difficult documents and create a page-level evaluation set this week. For each question, record the correct page and whether the answer is in prose, a table, a chart, or a footnote. Test retrieval separately from answer generation, then require the interface to open the cited page image. This exposes whether a system is finding evidence or merely producing plausible summaries.
Reader takeaway
For document assistants used where a wrong number has real cost, citations must be an operational feature rather than decoration. Cohere’s AWS example shows a concrete pattern: retrieve at page level, preserve layout when extraction is unreliable, select evidence before generation, and make every claim easy for a person to inspect.