Quick Research

The QuickResearchTool performs parallel web research across multiple queries. For each query, it searches the web using Tavily, crawls result pages with Crawl4AI, and summarizes page content via any AG2-supported LLM.

This is a lightweight alternative to DeepResearchTool for quick fact-finding across multiple topics in parallel.

How It Works#

For each query provided, the tool:

Searches the web using Tavily to get the top results
Crawls each result URL using Crawl4AI to extract page content
Summarizes each page using the llm_config LLM — raw page content (often 50,000+ tokens) is condensed into a short summary so it fits within the calling agent's context window

The tool returns structured JSON containing per-page summaries. The calling agent then synthesizes these into a final cohesive answer.

All queries are researched in parallel for fast results.

Package Installation#

Install AG2 with the quick-research extra (and openai for the example below):

pip install -U "ag2[openai,quick-research]"

Note: autogen and ag2 are aliases for the same PyPI package:
pip install -U "autogen[openai,quick-research]"

Environment Setup#

A Tavily API key is required for web search (free tier available):

export TAVILY_API_KEY="your_tavily_api_key"
export OPENAI_API_KEY="your_openai_api_key"

Implementation#

Imports#

import os
from autogen import AssistantAgent, LLMConfig
from autogen.tools.experimental import QuickResearchTool

Agent Configuration#

llm_config = LLMConfig({"model": "gpt-4o-mini", "api_type": "openai"})

assistant = AssistantAgent(
    name="researcher",
    system_message=(
        "You are a research assistant. Use the quick_research tool to look up information, "
        "then synthesize the results into a clear answer."
    ),
    llm_config=llm_config,
)

Tool Setup#

research_tool = QuickResearchTool(
    llm_config=llm_config,
    tavily_api_key=os.getenv("TAVILY_API_KEY"),
    num_results_per_query=2,
)

Usage Example#

assistant.run(
    message="Research the latest developments in quantum computing and summarize your findings.",
    tools=[research_tool],
    max_turns=2,
).process()

Parameters#

QuickResearchTool accepts the following initialization parameters:

Parameter	Type	Default	Description
`llm_config`	`LLMConfig`	required	LLM used internally to summarize each crawled page before returning results. Supports any AG2-compatible provider. This can be a different (smaller/cheaper) model than the one powering your agent.
`tavily_api_key`	`str \\| None`	`None`	Tavily API key. Falls back to `TAVILY_API_KEY` env var.
`num_results_per_query`	`int`	`3`	Number of search results to crawl per query.

At call time, the agent provides:

Parameter	Type	Default	Description
`queries`	`list[str]`	required	List of search queries to research (max 5).
`chunk_prompt`	`str`	`"Summarise the chunk, preserving all facts."`	Prompt for summarizing individual text chunks.
`merger_prompt`	`str`	`"Merge the partial summaries into one coherent overview."`	Prompt for merging chunk summaries.
`full_prompt`	`str`	`"Provide a concise but complete summary."`	Prompt for summarizing short texts in one pass.

Output#

The tool returns a JSON string containing a list of objects, one per query:

[
  {
    "query": "quantum computing breakthroughs",
    "sources": [
      {
        "title": "Page Title",
        "url": "https://example.com/article",
        "summary": "A concise summary of the page content..."
      }
    ]
  }
]