From the source
Cohere published a guide for building a financial research assistant on AWS that searches filings and answers questions with page-level citations.
- The design combines Cohere Embed for search, Rerank for ordering results, and Parse for turning hard pages into clean text, with
Anthropic Claude on
Amazon Bedrock writing the answers.
- Cohere says the problem is that the number an analyst needs usually sits in a table, footnote or chart, and that converting a PDF to plain text loses row and column relationships, reorders multi-column pages and drops charts.
- The guide recommends Embed Pro for high-quality document indexing and Embed Fast for low-latency queries, with configurable embedding dimensions and quantization to balance accuracy and storage.
- Rerank Pro reorders retrieved pages, and Context Focus marks the passages that answer the question to cut the text sent to the answer model.
- A CloudFormation-based demo lets teams deploy the stack in their own AWS account and compare results with and without Rerank.
