Multi-Agent Financial Analysis System for Equity Research

captide · Finance · 2025 · source Agentic: yes · promotional bias: high · 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: Equity research

Captide needed to let analysts define flexible financial-analysis tasks in natural language instead of manually searching regulatory filings and investor-relations documents through fixed legacy schemas.

Constraints:

Financial analysis needs precision and schema-compliant outputs.
Relevant information lives across filings and investor-relations materials.
Users may request custom table schemas and company-specific analysis.
Production agents need monitoring, feedback, error tracking, and cost visibility.

Architecture

Captide built a LangGraph Platform agent workflow where users specify analysis tasks in natural language and agents orchestrate retrieval and processing over regulatory filings and investor-relations documents. LangGraph supports parallel document queries and structured output through trustcall, while LangSmith captures traces, response times, error rates, costs, and user feedback.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  natural-language-interface["Natural Language Interface"] --> langgraph-agent-workflow["LangGraph Agent Workflow"]
  langgraph-agent-workflow --> financial-document-vector-stores["Financial Document Vector Stores"]
  langgraph-agent-workflow --> structured-output-layer["Structured Output Layer"]
  langgraph-agent-workflow --> langsmith["LangSmith"]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Natural Language Interface`	Captures analyst-defined research tasks.	`LangGraph Agent Workflow`
`LangGraph Agent Workflow`	Orchestrates retrieval, processing, grading, and structured output generation.	`Financial Document Vector Stores`, `Structured Output Layer`, `LangSmith`
`Financial Document Vector Stores`	Index regulatory filings and investor-relations documents for ticker-specific retrieval.	`LangGraph Agent Workflow`
`Structured Output Layer`	Uses trustcall and JSON schemas to enforce table and metric output contracts.	—
`LangSmith`	Provides tracing, monitoring, cost visibility, and user-feedback collection.	`LangGraph Agent Workflow`

Data flow

An analyst states a research task in natural language. Captide's agents run ticker-specific vector-store retrieval over financial documents, process relevant chunks in parallel, grade or structure the extracted information, and return schema-compliant outputs such as custom tables. LangSmith traces the workflow for monitoring and feedback.

“Once a user defines their analysis requirements, Captide's agents take over the entire data retrieval and processing pipeline.”summary, stated

“multiple agents work simultaneously to execute queries, retrieve documents, and grade document chunks.”summary, stated

Why this architecture

Captide used agentic workflows because equity research requires flexible, user-defined analysis over unstructured documents, not fixed-schema lookup. Parallel graph execution reduced latency, structured output protected downstream reliability, and LangSmith supplied monitoring and feedback loops for production use.

“the fixed schemas of legacy financial data platforms that cannot accommodate company-specific or customized analysis requirements.”summary, stated

“A critical requirement for financial applications is the ability to produce consistent, schema-compliant outputs.”summary, stated

Models used: unspecified LangGraph-hosted LLM agents

Scale

The platform reportedly compresses research that previously took days into seconds.
“reportedly compressing research that previously took days into seconds while maintaining the precision required for financial industry standards.”summary, stated

Alternatives considered (roads not taken)

Use legacy fixed-schema financial data platforms

Rejected because: Fixed schemas cannot handle company-specific or custom analysis requirements.

“fixed schemas of legacy financial data platforms that cannot accommodate company-specific or customized analysis requirements.”summary, stated

Run document queries sequentially

Rejected because: Parallel execution minimizes latency without complicating the codebase with async functions.

“This parallel approach is emphasized as a way to minimize latency without complicating the codebase with asynchronous functions.”summary, stated

Patterns

`rag-retrieval` (knowledge)

Captide retrieves relevant filing and investor-relations chunks from ticker-specific vector stores.

“the text mentions "ticker-specific vector store queries" as part of the agent workflow.”summary, stated

`structured-output-contract` (reliability)

The platform uses predefined JSON schemas and trustcall for custom table outputs.

“Captide uses LangGraph's trustcall Python library to ensure that outputs adhere strictly to predefined JSON schemas.”summary, stated

`observability-tracing` (operations)

LangSmith captures traces, latency, errors, costs, and feedback.

“The platform captures traces of agent workflows that include response times, error rates, and operational costs.”summary, stated

Tradeoffs

Use parallel LangGraph execution

Chose Concurrent document queries and chunk grading over Sequential document processing.

Rationale: Parallel execution reduces latency for research workflows over many filings.

“multiple agents work simultaneously to execute queries, retrieve documents, and grade document chunks.”summary, stated

Enforce structured outputs

Chose trustcall and JSON schemas over Free-form financial analysis text only.

Rationale: Financial outputs often feed tables and downstream systems that need consistent schemas.

“unreliable output formats can break downstream systems.”summary, stated

Gotchas (what bit them)

Financial users need custom schemas rather than fixed vendor fields.

Root cause: Company-specific metrics and ad hoc analysis do not fit legacy fixed schemas.

Resolution: Captide allows natural-language analysis requests and structured custom outputs.

“users request table outputs with custom schemas to structure metrics found across multiple documents.”summary, stated

Production financial agents require cost and error observability.

Root cause: Financial workflows are latency- and precision-sensitive, and excessive API calls can become expensive.

Resolution: LangSmith traces capture response time, error rate, operational cost, and user feedback.

“Understanding the cost of LLM operations is particularly important in financial applications where margins matter and excessive API calls can quickly become expensive.”summary, stated

Evaluation

Captide uses LangSmith tracing and direct user feedback, but the public case does not give detailed quantitative accuracy metrics.

Metrics: response times · error rates · operational costs · thumbs-up/thumbs-down feedback

“Captide has integrated thumbs-up and thumbs-down feedback mechanisms directly into their platform.”summary, stated

Outcomes

Captide deployed production-ready financial analysis agents on LangGraph Platform.
The platform lets analysts define custom research tasks in natural language.
Research workflows reportedly moved from days to seconds.

The source claims days-to-seconds compression but does not disclose volume.

“they deployed production-ready agents on LangGraph Platform.”summary, stated