Skip to main content

View Source Code

Check out the full source code for this example on GitHub.
 This tutorial demonstrates how to build a Deep Research Agent using Tensorlake and the OpenAI Agents SDK. This application orchestrates multiple agents to plan, search, and write comprehensive research reports on any given topic.

​
Overview

The Deep Research Agent consists of three specialized agents that work together in a pipeline:
  1. Planner Agent: Breaks down the user’s research topic into specific search queries and steps.
  2. Search Agent: Executes the planned search queries in parallel, retrieving and summarizing relevant information from the web.
  3. Writer Agent: Synthesizes the gathered information into a structured, comprehensive markdown report.
Each agent runs as an isolated, serverless function on Tensorlake, ensuring scalability and fault tolerance.

​
Prerequisites

  • Python 3.11+
  • Tensorlake Account and CLI installed.
  • OpenAI API Key

​
Project Structure

Your project should look like this: 
deep-research/
├── app.py          # Main application logic and Tensorlake functions
├── models.py       # Pydantic data models for structured inputs/outputs
├── prompts.py      # System prompts for the agents
└── requirements.txt

​
Implementation

​
1. Define Data Models (models.py)

First, we define the data structures that our agents will use to communicate. This ensures type safety and clear interfaces between the agents. 
from pydantic import BaseModel, Field
from typing import List

class SearchQuery(BaseModel):
    query: str = Field(..., description="A specific search query to execute.")
    rationale: str = Field(..., description="Why this query is important.")

class ResearchPlan(BaseModel):
    topic: str
    search_queries: List[SearchQuery]
    
class SearchResult(BaseModel):
    url: str
    title: str
    content: str
    summary: str

class ResearchReport(BaseModel):
    topic: str
    markdown_content: str
    references: List[str]

​
2. Create the Agents (app.py)

In app.py, we define our Tensorlake functions. Each function represents a stage in the pipeline and utilizes an OpenAI agent. 
import os
from typing import List

from openai import OpenAI
from tensorlake import application, function, Image
from pydantic import BaseModel

from models import ResearchPlan, SearchResult, ResearchReport
# Import your prompts here
# from prompts import PLANNER_PROMPT, SEARCH_PROMPT, WRITER_PROMPT

# Define the runtime image
image = Image(name="deep-research-agent").run("pip install openai tensorlake pydantic")

@application()
@function(image=image, secrets=["OPENAI_API_KEY"])
def deep_research_pipeline(topic: str) -> ResearchReport:
    """
    Orchestrates the deep research pipeline.
    """
    print(f"Starting deep research on: {topic}")
    
    # Phase 1: Planning
    plan = create_research_plan(topic)
    print(f"Plan created with {len(plan.search_queries)} queries.")
    
    # Phase 2: Searching (Parallel Execution)
    # We map the search function over the queries to run them in parallel
    search_results = execute_search.map(plan.search_queries)
    print(f"Completed {len(search_results)} searches.")
    
    # Phase 3: Writing
    report = write_report(topic, search_results)
    print("Report generation complete.")
    
    return report

@function(image=image, secrets=["OPENAI_API_KEY"])
def create_research_plan(topic: str) -> ResearchPlan:
    client = OpenAI()
    completion = client.beta.chat.completions.parse(
        model="gpt-4o",
        messages=[
            {"role": "system", "content": "You are an expert research planner."},
            {"role": "user", "content": f"Create a research plan for: {topic}"}
        ],
        response_format=ResearchPlan
    )
    return completion.choices[0].message.parsed

@function(image=image, secrets=["OPENAI_API_KEY"])
def execute_search(query_obj) -> SearchResult:
    # In a real implementation, you would use a search tool or API here.
    # For this example, we'll simulate a search result.
    # You could use tools like Tavily, Serper, or a custom scraper.
    
    client = OpenAI()
    # Simulate processing the query
    summary = f"Simulated search results for: {query_obj.query}"
    
    return SearchResult(
        url="https://example.com",
        title=f"Results for {query_obj.query}",
        content="Full content would go here...",
        summary=summary
    )

@function(image=image, secrets=["OPENAI_API_KEY"])
def write_report(topic: str, results: List[SearchResult]) -> ResearchReport:
    client = OpenAI()
    
    # Compile context from search results
    context = "

".join([f"Source: {r.url}
Summary: {r.summary}" for r in results])
    
    completion = client.beta.chat.completions.parse(
        model="gpt-4o",
        messages=[
            {"role": "system", "content": "You are an expert research writer. Write a comprehensive report based on the provided context."},
            {"role": "user", "content": f"Topic: {topic}

Context:
{context}"}
        ],
        response_format=ResearchReport
    )
    return completion.choices[0].message.parsed

​
Running Locally

To test your pipeline locally, add this code block to the end of app.py and run it with python. 
if __name__ == "__main__":
    from tensorlake.applications import run_local_application
    
    # Run the application locally
    run_local_application(deep_research_pipeline, "The Future of Quantum Computing")
Then run the script: 
export OPENAI_API_KEY=your_key_here
python app.py

​
Deploying to Tensorlake

When you’re ready to deploy, use the tensorlake deploy command. 
tensorlake secrets set OPENAI_API_KEY=your_key_here
tensorlake deploy app.py
Your deep research agent is now live and scalable! You can invoke it via the provided HTTP endpoint or the Tensorlake SDK.