graph TD %%{init: {'theme': 'mc','layout': 'elk'}}%% ChatInput-im722[<div><img alt="logo" src="/_astro/messages-square.BaSDmT6g.svg" style="height: 20px !important;width: 20px !important"/></div>Chat Input] style ChatInput-im722 stroke:#a170ff ChatOutput-7h4uy[<div><img alt="logo" src="/_astro/messages-square.BaSDmT6g.svg" style="height: 20px !important;width: 20px !important"/></div>Chat Output] style ChatOutput-7h4uy stroke:#a170ff OpenAIModel-xyjkl[<div><img alt="logo" src="/_astro/openAI.BhmuxEs3.svg" style="height: 20px !important;width: 20px !important"/></div>OpenAI] style OpenAIModel-xyjkl stroke:#a170ff DeepResearchAgent-af1xa[Deep Research Agent] style DeepResearchAgent-af1xa stroke:#a170ff OpenAIWebSearchToolComponent-mezqb[<div><img alt="logo" src="/_astro/openAI.BhmuxEs3.svg" style="height: 20px !important;width: 20px !important"/></div>OpenAI Built-in tools] style OpenAIWebSearchToolComponent-mezqb stroke:#a170ff ChatInput-im722 -.- DeepResearchAgent-af1xa linkStyle 0 stroke:#a170ff DeepResearchAgent-af1xa -.- ChatOutput-7h4uy linkStyle 1 stroke:#a170ff OpenAIModel-xyjkl -.- DeepResearchAgent-af1xa linkStyle 2 stroke:#a170ff OpenAIWebSearchToolComponent-mezqb -.- DeepResearchAgent-af1xa linkStyle 3 stroke:#a170ff

Deep Research Agent Workflow Documentation

🧩 Overview

The Deep Research Agent automates complex information gathering by following a four‑phase protocol: Plan, Act, Process, Synthesize. It receives a user query, constructs a detailed research plan, executes searches and analyses through a built‑in web‑search tool, and returns a concise, structured report. This workflow streamlines deep‑dive investigations, reducing manual effort and ensuring consistent, high‑quality outputs.

⚙️ Main Features

Automated Planning – The agent formulates a step‑by‑step research roadmap based on the user’s initial prompt.
Human‑in‑the‑Loop Validation – Before execution, the plan is presented for user approval or modification.
Integrated Web Search – A built‑in web‑search tool retrieves up‑to‑date information from the internet.
Iterative Refinement – The agent continually revises its plan and queries until all required data are gathered.
Structured Reporting – Final outputs follow a defined schema, ensuring consistency across responses.
Streaming Capability – Optional streaming lets users see the agent’s reasoning in real time.

🔄 Workflow Steps

Component Name	Role in the Workflow	Key Inputs	Key Outputs
Chat Input	Receives the user’s query and optional files.	User text, session ID, optional file attachments.	Message object containing the prompt and metadata.
OpenAI Model	Provides the large‑language‑model engine for planning and content generation.	Prompt, system message, model configuration (temperature, max tokens, etc.).	Language‑model instance used by the agent.
OpenAI Web Search Tool	Supplies a web‑search capability to fetch real‑time data.	Search query, selected model, tool configuration.	Tool instance that the agent can invoke during execution.
Deep Research Agent	Orchestrates the full research cycle: plan, validate, execute, synthesize.	User prompt, LLM instance, web‑search tool, optional memory and streaming settings.	Structured response message ready for display.
Chat Output	Presents the final answer to the user.	Response message from the agent.	Rendered chat message in the playground UI.

Note: The Label component only displays static documentation and does not participate in the data flow.

🧠 Notes

The agent relies on an OpenAI API key; the key must be configured in the OpenAI Model component.
Web‑search results depend on the selected OpenAI model (e.g., gpt‑4o‑mini).
Token limits of the chosen LLM constrain the depth of the plan and the detail of the final report.
Human‑in‑the‑Loop approval is mandatory before any external tools are executed; disabling this feature removes the safety check.
Streaming is optional and may increase latency; it is enabled only when the user sets the corresponding flag.
Structured output requires a JSON schema; if the agent’s response does not match the schema, the workflow will fail.
Memory is optional; enabling it allows the agent to preserve context across multiple interactions but increases resource usage.
The workflow is stateless between runs unless memory is activated; each execution starts with a fresh agent instance.