AI Assistant for Financial Data Discovery and Business Intelligence

amazon_finance · Finance · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Financial data discovery

Amazon Finance analysts spent too much time finding financial data, reconciling systems, and recovering institutional knowledge across documents and legacy systems.

Constraints:

• enterprise security
• semantic search
• financial accuracy
• global operations

Architecture

Amazon Finance built a RAG assistant around a Streamlit chat UI, Amazon Kendra Enterprise Edition, and Claude 3 Sonnet in Amazon Bedrock. Analysts ask natural-language finance questions; Kendra retrieves relevant paragraphs and document references from indexed financial knowledge; Bedrock combines that context with a prompt template for Claude; the UI returns the answer, source context, feedback controls, and session history. Kendra was chosen over a simpler company-search option because the finance domain needed relevance tuning, metadata customization, synonyms, and pre-generation monitoring.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  streamlit-chat-ui["Streamlit chat UI"]
  amazon-kendra-retriever["Amazon Kendra retriever"]
  kendra-index["Kendra index"]
  amazon-bedrock-rag-solution["Amazon Bedrock RAG solution"]
  claude-3-sonnet["Claude 3 Sonnet"]
  s3-document-buckets["S3 document buckets"]
  feedback-store["feedback store"]
  streamlit-chat-ui --> amazon-kendra-retriever
  amazon-kendra-retriever --> kendra-index
  kendra-index --> s3-document-buckets
  amazon-kendra-retriever --> amazon-bedrock-rag-solution
  amazon-bedrock-rag-solution --> claude-3-sonnet
  amazon-bedrock-rag-solution --> streamlit-chat-ui
  streamlit-chat-ui --> feedback-store

Components

Data flow

An authenticated analyst asks a question in the Streamlit UI. The app sends the query to Amazon Kendra, which retrieves relevant paragraphs and document references from its index. The Bedrock RAG layer sends that retrieved context and prompt template to Claude 3 Sonnet, returns the generated answer to Streamlit, and records user feedback separately while Kendra continues indexing S3 documents.

Why this architecture

The architecture targets finance search problems that keyword tools handled poorly: analysts needed natural-language exploration, contextual relationships, preserved institutional knowledge, and controlled retrieval before generation. Kendra supplied semantic retrieval plus tuning and metadata control, while Bedrock/Claude supplied answer synthesis grounded in verified organizational knowledge.

Models used: Anthropic Claude 3 Sonnet · Amazon Kendra Enterprise Edition

Scale

• Benchmarking with analysts found a 30% reduction in search time and an 80% improvement in search-result accuracy.
  “During our benchmarking tests with analysts, we found that this solution dramatically reduced search time by 30% because analysts can now perform natural language search, and it improved the accuracy of search results by 80%.”article, stated
• The quantitative test set covered more than 50 typical analyst business queries.
  “We focused primarily on precision and recall testing, creating a diverse test set of over 50 business queries that represented typical use cases our analysts encounter.”article, stated
• Knowledge search reached 83% precision and 74% recall before metadata enrichment.
  “The knowledge search capability demonstrated initial rates of 83% precision and 74% recall without performing metadata enrichment on data sources.”article, stated
• Average time to find information fell from 45-60 minutes to 5-10 minutes, and 92% of analysts preferred the new system.
  “the new system reduced the average time to find relevant information from 45–60 minutes to just 5–10 minutes—an 85% improvement in efficiency. User satisfaction surveys indicate that 92% of analysts prefer the new system over traditional search methods”article, stated

Alternatives considered (roads not taken)

Traditional keyword search and rigid query tooling

Rejected because: Those tools missed contextual finance relationships and limited dynamic exploration.

Amazon Q Business as the retrieval layer

Rejected because: Kendra gave the team stronger retrieval accuracy, search-parameter control, custom document attributes, and custom synonyms for the finance domain.

Patterns

rag-retrieval (business-intelligence)

The assistant retrieves financial documents and context before generating answers.

hybrid-search (retrieval)

Amazon Kendra combines keyword, semantic, and vector search without manual implementation.

Tradeoffs

Search backend choice.

Chose Amazon Kendra Enterprise Edition over custom OpenSearch implementation.

Rationale: Kendra reduced manual search-configuration work through built-in enterprise connectors and NLU.

Gotchas (what bit them)

Traditional keyword search missed contextual relationships in financial data.

Root cause: Financial knowledge lived in disparate sources with rigid query structures and siloed historical decisions.

Resolution: Use enterprise search plus RAG to preserve institutional knowledge and support natural language discovery.

Evaluation

Amazon Finance measured search-time reduction, accuracy, precision, and faithfulness.

Metrics: 30% search-time reduction · 80% search accuracy improvement · 83% precision · 88% faithfulness

Outcomes

• Information discovery reportedly fell from 45-60 minutes to 5-10 minutes.

RAG assistant rather than clearly autonomous agent; marked non-agentic.