Multi-Agent Financial Research and Question Answering System

yahoo!_finance · Finance · 2024 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (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 research question answering

Yahoo! Finance needed to help retail users ask complex financial questions across structured market data, SEC filings, news, and APIs while maintaining accuracy, compliance boundaries, and production-scale latency/cost.

Constraints:

• Yahoo! Finance serves over 150 million users globally.
• The system has to handle structured and unstructured financial data with temporal fiscal-year complexity.
• Financial answers must avoid giving advice while remaining accurate and timely.
• Tool overload in a single agent causes confusion and prompt bloat.

Architecture

Yahoo! Finance uses a supervisor-subagent architecture on Amazon Bedrock Agent Core. Specialized agents handle structured financial data, insider transactions, SEC filings through RAG, and news through RAG plus restricted web search. The system combines ReAct-style adaptive reasoning, ReWOO-style parallel planning, and Reflexion-style self-review, with hybrid human/AI evaluation for quality.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  supervisor-agent["Supervisor Agent"] --> financial-data-agent["Financial Data Agent"]
  supervisor-agent --> insider-transaction-agent["Insider Transaction Agent"]
  supervisor-agent --> sec-filings-agent["SEC Filings Agent"]
  supervisor-agent --> news-agent["News Agent"]
  financial-data-agent --> financial-apis["Financial APIs"]
  insider-transaction-agent --> financial-apis
  sec-filings-agent --> sec-vector-knowledge-base["SEC Vector Knowledge Base"]
  news-agent --> news-knowledge-base["News Knowledge Base"]

Components

Data flow

A user asks a financial research question. The supervisor decomposes and routes the request to specialized subagents for prices, fundamentals, filings, news, or APIs. Structured agents call internal/partner APIs, RAG agents retrieve SEC or news evidence, and the supervisor synthesizes a compliant answer while maintaining session context.

Why this architecture

Yahoo chose supervisor-subagents because a single broad agent became error-prone when overloaded with overlapping tools and documentation. Separating structured-data, filings, news, and insider-transaction capabilities reduced tool confusion and let agents use reasoning patterns suited to financial research.

Models used: Amazon Bedrock Agent Core models

Scale

• Yahoo! Finance serves over 150 million users globally.
  “Yahoo! Finance serves over 150 million users globally as the leading financial news and information platform”summary, stated
• The source cites 4700 SEC filings and 3000 news articles daily.
  “The system must process thousands of SEC filings daily (4,700 filings), thousands of news articles (3,000 per day)”summary, stated
• The system is reported at 5-50 seconds and 2-5 cents per query.
  “while processing queries in 5-50 seconds at costs of 2-5 cents per query”summary, stated
• The system supports 100+ concurrent users.
  “demonstrating production viability at scale with support for 100+ concurrent users.”summary, stated

Alternatives considered (roads not taken)

Use one broad agent with all tools

Rejected because: Too many tools and overlapping definitions caused confusion and prompt bloat.

Use only ReAct-style sequential tool use

Rejected because: Financial research benefits from upfront planning, parallel data gathering, and self-reflection.

Patterns

multi-agent-orchestration (architecture)

A supervisor coordinates specialized agents for structured data, filings, news, and insider transactions.

rag-retrieval (knowledge)

SEC filings and news agents use specialized vector knowledge bases.

tool-use-orchestration (execution)

Structured-data agents call internal and partner APIs for financial data.

Tradeoffs

Use supervisor-subagent specialization

Chose Focused financial agents over A single all-purpose financial agent.

Rationale: Specialization reduces tool confusion and lets agents excel in focused domains.

Combine reasoning patterns

Chose ReAct, ReWOO, and Reflexion-style behaviors over One fixed reasoning loop.

Rationale: Financial research needs adaptive tool use, parallel gathering, and self-review.

Gotchas (what bit them)

A single agent became too broad for financial research.

Root cause: Many overlapping tools and API docs caused selection confusion and prompt bloat.

Resolution: Yahoo split the system into focused subagents under a supervisor.

Retail finance QA must avoid financial-advice boundary violations.

Root cause: The system answers investment research questions for a high-stakes financial domain.

Resolution: The stated goal includes maintaining accuracy and avoiding financial advice.

Evaluation

Yahoo uses a hybrid evaluation approach combining human and AI judges, with reported accuracy, coverage, latency, cost, and concurrency metrics.

Metrics: 5-50 seconds per query · 2-5 cents per query · 100+ concurrent users · human and AI judges

Outcomes

• Yahoo built a production-scale multi-agent financial QA system.
• The architecture made heterogeneous data accessible through specialized agents.
• Reported production viability includes 5-50 second latency, 2-5 cent query costs, and 100+ concurrent users.

The case cites 150 million users, 4700 filings/day, 3000 news articles/day, 5-50 second latency, and 100+ concurrent users.