Cohere Demonstrates Filing Research Assistant on 928 Pages, With Explicit Production Gaps

On October 29, 2026, Cohere published a technical guide to a financial-filings research assistant built with its models and AWS services, connecting retrieval, reranking, page images and answers with page-level citations. Its deployable demonstration uses a fixed set of three public fiscal-2024 annual reports totaling 928 pages, and the authors explicitly call it a starting point rather than a production system. The design discusses Embed Pro, Embed Fast, Rerank Pro and Parse, but the actual demo intentionally uses a different subset and configuration; an architectural possibility should not be reported as a feature already present in the demo. This site's issue date is November 2.

The task begins with finding a precise figure in a document hundreds of pages long. A number may sit in a table, footnote or chart, where text-only search can lose row labels, the year or the unit. Cohere's architecture indexes pages with Embed, retrieves likely pages for a question, optionally uses Rerank to reorder them and can use Context Focus to identify relevant passages. Claude on Amazon Bedrock then writes an answer from retrieved text and page images, with a page citation and a content type such as table, chart, text or footnote for each claim. A citation is a useful user interface for checking the evidence, not a guarantee that the model interpreted the financial statement correctly. The user still has to inspect the underlying page.

The guide is unusually explicit about the gap between proposed design and simplified demo. Its preferred architecture uses Embed Pro when indexing and Embed Fast for interactive questions, with Rerank Pro and Parse added when their quality benefit justifies latency and cost. The demo instead uses Embed Pro for both indexing and queries and Rerank Fast to keep GPU needs lower. It indexes text already extracted by the benchmark dataset. It does not ingest arbitrary uploaded PDFs, nor does it actually run complex scanned pages through Parse. Saying that this particular demo automatically parses any scanned filing would promote a future architectural option into a current capability. A company's own documents may differ sharply from three public reports; a representative test set with known answers should come first.

Vector storage presents another reproducible tradeoff. Cohere describes a 1,024-dimension int8 vector that occupies about one kilobyte of raw storage per page, compared with roughly eight kilobytes for a 2,048-dimension float32 vector. That eightfold comparison concerns raw vector bytes; it excludes index overhead, metadata, replicas and OpenSearch charges, and it does not prove better retrieval quality. Query and index embeddings must share dimension and format; changing them later generally means re-embedding the pages. For mostly digital reports with clean text, page text is an economical retrieval input. For scans, dense tables and figures, parsing or image retrieval may matter more. A useful comparison measures page-level recall, citation correctness, latency and total cost for each path.

The demo's security and operational scope must not be hidden. It includes one user, an OpenSearch Serverless collection with public network access protected by a data access policy, and no Bedrock Guardrails. These are simplifications for evaluation, not a recommended production posture. Cohere suggests a VPC endpoint and private OpenSearch access, an organizational identity provider, per-user or per-team document permissions, monitoring, alarms, endpoint autoscaling and a budget alarm before production. Its real-time SageMaker model endpoints bill for every hour they exist, including idle time. A trial therefore needs a cleanup step from the repository instructions, not merely closing the browser window.

An engineer can begin in an isolated AWS account by checking Marketplace subscriptions, Bedrock model access and the region, then deploying the public fixed dataset using the repository README. Test questions with known answers should verify which page is retrieved, whether Rerank changes the order beneficially, whether Context Focus drops a table heading or unit, and whether each Claude citation points to evidence that supports the exact claim. Switch page images and reranking on and off while recording tokens, latency and cost per question. Before introducing internal filings, add private network access, identity, encryption, data classification and a cost budget, and specify how the assistant will say that the source pages do not contain an answer. A demo score cannot replace those controls.

The useful contribution is a set of testable choices for retrieval, table context and checkable citations, with a candid account of how the sample differs from the intended architecture. Claims that the demo is ready for production, already parses arbitrary PDFs or provides private networking by default exceed the source; data, access and costs need verification first.