Deep research agent using Vercel AI SDK

This example project is derived from the brilliant deep research guide by Nico Albanese.

Overview

This full-stack Next.js project is an intelligent deep research agent that autonomously conducts multi-layered web research, generates a comprehensive structured report, converts it to a PDF, and uploads it to cloud storage — all with real-time progress updates on the frontend.

Tech stack:

Next.js — web application
Vercel AI SDK — AI model integration and structured generation
Trigger.dev — task orchestration, execution, and real-time progress
OpenAI GPT-4o — query generation, content analysis, report writing
Exa API — semantic web search with live crawling
LibreOffice — HTML-to-PDF conversion
Cloudflare R2 — stores the generated PDF reports

GitHub repo

View the deep research agent repo

Fork this repo to use it as a starting point for your own research agent.

How the agent works

Task orchestration

Three Trigger.dev tasks are chained together using triggerAndWait, each handling a distinct phase of the pipeline:

deepResearchOrchestrator — coordinates the entire research workflow
generateReport — converts accumulated research into a structured HTML report using GPT-4o
generatePdfAndUpload — converts the HTML to PDF with LibreOffice and uploads to R2

src/trigger/deepResearch.ts

import { task, metadata } from "@trigger.dev/sdk";
import { generateReport } from "./generateReport";
import { generatePdfAndUpload } from "./generatePdfAndUpload";

export const deepResearchOrchestrator = task({
  id: "deep-research",
  maxDuration: 300, // 5 minutes
  run: async (payload: { query: string; depth?: number; breadth?: number }) => {
    const depth = payload.depth ?? 2;
    const breadth = payload.breadth ?? 2;

    metadata.set("status", { progress: 0, label: "Starting research..." });

    // Recursively research the query
    const { learnings, sources } = await deepResearch({
      query: payload.query,
      depth,
      breadth,
    });

    metadata.set("status", { progress: 75, label: "Generating report..." });

    // Chain to the report generation task
    const reportResult = await generateReport.triggerAndWait({
      query: payload.query,
      learnings,
      sources,
    });

    if (!reportResult.ok) throw new Error("Report generation failed");

    metadata.set("status", { progress: 90, label: "Converting to PDF..." });

    // Chain to the PDF generation task
    const pdfResult = await generatePdfAndUpload.triggerAndWait({
      html: reportResult.output.html,
      filename: `research-${Date.now()}.pdf`,
    });

    if (!pdfResult.ok) throw new Error("PDF generation failed");

    metadata.set("status", { progress: 100, label: "Done!" });

    return { pdfUrl: pdfResult.output.url, sources };
  },
});

Recursive research algorithm

The core research logic uses a recursive depth-first approach. At each level the agent:

Generates multiple search queries from the current question using GPT-4o
Searches the web via the Exa API
Evaluates the relevance of each result
Extracts key learnings and follow-up questions
Recurses on the follow-up questions at the next depth level

Level 0: "AI safety in autonomous vehicles"
├── Level 1 (breadth=2):
│   ├── "Machine learning safety protocols in self-driving cars"
│   │   └── Follow-up: "How do neural networks handle edge cases?"
│   └── "Regulatory frameworks for autonomous vehicle testing"
│       └── Follow-up: "What are current safety certification requirements?"
└── Level 2 (breadth=1, depth limit):
    ├── "Neural network edge case handling" → STOP
    └── "Safety certification requirements" → STOP

Real-time progress with Trigger.dev Realtime

The frontend uses the useRealtimeTaskTrigger hook to both trigger the task and subscribe to progress updates without polling:

src/components/DeepResearchAgent.tsx

import { useRealtimeTaskTrigger } from "@trigger.dev/react-hooks";
import type { deepResearchOrchestrator } from "../trigger/deepResearch";

export function DeepResearchAgent({ accessToken }: { accessToken: string }) {
  const triggerInstance = useRealtimeTaskTrigger<typeof deepResearchOrchestrator>(
    "deep-research",
    { accessToken }
  );

  const status = triggerInstance.run?.metadata?.status;
  const progress = status?.progress ?? 0;
  const label = status?.label ?? "";

  async function handleSubmit(query: string) {
    await triggerInstance.submit({ query, depth: 2, breadth: 2 });
  }

  return (
    <div>
      <button onClick={() => handleSubmit("Your research question here")}>
        Start research
      </button>
      {triggerInstance.run && (
        <div>
          <p>{label}</p>
          <progress value={progress} max={100} />
        </div>
      )}
    </div>
  );
}

As research progresses, the task updates its metadata and the frontend receives each update automatically:

src/trigger/deepResearch.ts

metadata.set("status", {
  progress: 25,
  label: `Searching the web for: "${query}"`,
});

Relevant code files

File	Description
`src/trigger/deepResearch.ts`	Core recursive research logic and orchestrator task
`src/trigger/generateReport.ts`	Converts research data into a structured HTML report
`src/trigger/generatePdfAndUpload.ts`	Converts HTML to PDF with LibreOffice and uploads to R2
`src/components/DeepResearchAgent.tsx`	Frontend form and real-time progress display

Learn more

Trigger.dev Realtime

Subscribe to runs and stream live progress to your frontend

React hooks

Trigger a task and subscribe to its updates from a React component

Run metadata

Store and broadcast arbitrary data from inside a running task

triggerAndWait

Chain tasks together and wait for each to complete before continuing

Guides

Example Projects

Example Tasks

Deep research agent using Vercel AI SDK

Overview

GitHub repo

View the deep research agent repo

How the agent works

Task orchestration

Recursive research algorithm

Real-time progress with Trigger.dev Realtime

Relevant code files

Learn more

Trigger.dev Realtime

React hooks

Run metadata

triggerAndWait

​Overview

​GitHub repo

View the deep research agent repo

​How the agent works

​Task orchestration

​Recursive research algorithm

​Real-time progress with Trigger.dev Realtime

​Relevant code files

​Learn more

Trigger.dev Realtime

React hooks

Run metadata

triggerAndWait

Overview

GitHub repo

How the agent works

Task orchestration

Recursive research algorithm

Real-time progress with Trigger.dev Realtime

Relevant code files

Learn more