Cut Search Time by 30% and Boost Accuracy 80% with Amazon Bedrock for Financial Data Retrieval

Problem Statement

Financial analysts in Amazon Finance needed to (1) reduce manual effort spent browsing multiple data catalogs, (2) consolidate fragmented historical and decision‑making data, and (3) obtain metrics quickly in a rapidly changing business environment. Keyword‑based search and rigid query structures failed to capture contextual relationships, causing duplicated analysis, inconsistent assumptions, and loss of institutional knowledge.

Solution Overview

An AI‑assistant was built using generative AI on Amazon Bedrock (Anthropic Claude 3 Sonnet) combined with Amazon Kendra’s enterprise‑grade semantic search. Analysts interact via natural‑language queries that flow through a Retrieval‑Augmented Generation (RAG) pipeline.

Core Components

Intelligent Retrieval : A vector database stores high‑dimensional embeddings of all financial documents. User queries are embedded and matched semantically, enabling keyword‑free search.

Enhanced Generation : Retrieved context is supplied to Claude 3 Sonnet with a prompt template that encourages concise, fact‑based answers and reduces hallucinations.

Prompt Template Example

"""H: Use the following pieces of context to provide a concise answer to the question at the end. If you don't know the answer, just say that you don't know.
<context>{context}</context>
Question: {question}
A:"""

Architecture

Amazon Bedrock – hosts Claude 3 Sonnet.

Amazon Kendra Enterprise – provides semantic, keyword, and vector search with built‑in NLU, support for 40+ document formats, and custom synonym handling.

Streamlit – Python UI for query entry and response display.

AWS infrastructure: Route 53, CloudFront, Lambda (authentication), Fargate/ECS (container runtime), S3 (feedback storage), and Kendra indexing.

Evaluation Framework

A test set of 50+ business queries was used to measure quantitative metrics (precision, recall, latency) and qualitative faithfulness.

Search‑time reduction: average latency dropped from 45‑60 minutes to 5‑10 minutes (≈85 % faster).

Overall accuracy improvement: +80 %.

Data‑discovery precision = 65 %, recall = 60 % (baseline manual search ≈35 %).

Document‑search precision = 83 %, recall = 74 % (keyword baseline 45‑50 %).

LLM‑as‑judge faithfulness: 70 % for data‑discovery, 88 % for knowledge‑retrieval.

92 % of surveyed analysts preferred the new system.

Key Benefits and Use Cases

Analysts can ask natural‑language questions such as “Where can I find the production‑efficiency metric?” and receive precise, context‑aware answers without knowing database schemas. Demonstrated scenarios include:

Rapid data‑source identification.

Extraction of specific financial metrics.

Interpretation of enterprise planning documents.

Limitations and Future Work

Current challenges are handling highly complex queries that require fine‑grained disambiguation and ensuring comprehensive coverage of newly added data sources. Ongoing work focuses on enriching metadata and expanding the knowledge base.

Conclusion

The AI‑assistant built on Amazon Bedrock, Claude 3 Sonnet, and Amazon Kendra shows that generative AI combined with enterprise semantic search can dramatically improve efficiency and accuracy for financial analysts, providing a blueprint for similar deployments in other domains.