Cohere has published a practical design for a financial research assistant that answers questions about company filings and cites the page behind every claim. The useful twist is not just finding a number in a long report, but making it quick to check where that number came from.
Cohere Watch analysis
What happened
Cohere’s financial research assistant guide combines Cohere models with AWS services. It uses Embed Pro to index filing pages, Embed Fast to search when a question comes in, Rerank to reorder promising results, and Anthropic Claude on Amazon Bedrock to write the answer. Cohere Parse can extract tables from pages with weak text or scanned content.
The proposed assistant attaches a page citation to each claim and identifies whether its source is a table, chart, body text, footnote or figure. Readers can open the page image to check it. Cohere also provides a CloudFormation-based demo for deploying and experimenting in an AWS account.
Our top picks
- Page-level citations
Each answer claim points to its filing page and the type of evidence used. - Separate models for indexing and live search
Embed Pro handles one-time indexing; Embed Fast handles incoming questions. - Reranking with Context Focus
Rerank can reorder candidate pages and identify the passages relevant to an answer. - Parsing difficult pages
Cohere Parse can turn scanned or table-heavy pages into structured text. - Storage and cost choices
The guide uses 1,024-dimension int8 vectors and explains how other sizes and formats trade storage against retrieval quality.
Why it matters
Financial answers often hide in tables, footnotes and charts, where converting a filing into plain text can scramble the evidence. This design combines searchable text with page images and makes the source visible, giving analysts a way to check an answer rather than simply accept a fluent paragraph.
It also shows how the components fit together: index pages once, retrieve and optionally rerank them for each question, then have Claude produce a cited response. Cohere’s guide describes a proposed implementation, not an independent evaluation of its accuracy. For financial work, that distinction is rather more than fine print.
Our read
The best feature here is the trail back to the document. A generated answer that shows its workings is more useful than one that merely sounds as if it has read the footnotes. Teams can use the guide and demo to explore the architecture, but should test retrieval and citations against their own filings before relying on the result.
What to watch
- Whether the demo and its cited answers work reliably across real filings.
- How retrieval quality changes with different vector sizes and reranking settings.
- Whether page-level citations make it easier for analysts to spot errors in practice.
Discussion spark: For financial AI, is a citation on every claim enough to make an answer trustworthy, or should analysts still check each source page themselves?
Sources and evidence
- Source update (29 October 2026, 21:02 UTC)
not affiliated with or endorsed by Cohere