Skip to main content

Search Agent

The Search Agent enables flows to retrieve information from external sources. It supports two primary search modes: live web search and semantic search over your private documents.

Providers​

The agent routes its execution based on the parameters provided in model_params:

  1. Google: Live web search using the Google Custom Search JSON API.
  2. Intellicloud: Semantic search over data indexed in the Intellicloud platform.

Parameters​

  • k: Number of search results to return (default: 5 for Google, 3 for Intellicloud).
  • as_text: (Google only) If True, returns a formatted string. If False, returns a structured list of results.

To use Google search, you need a Google API Key and a Custom Search Engine ID (CX).

import asyncio
from intelli.flow.agents.agent import Agent
from intelli.flow.types import AgentTypes
from intelli.flow.input.task_input import TextTaskInput
from intelli.flow.tasks.task import Task
from intelli.flow.sequence_flow import SequenceFlow

# 1. Define the Google Search Agent
search_agent = Agent(
    agent_type=AgentTypes.SEARCH.value,
    provider="google",
    mission="find latest news about AI",
    model_params={
        "google_api_key": "YOUR_GOOGLE_API_KEY",
        "google_cse_id": "YOUR_CSE_ID",
        "k": 3,
        "as_text": True
    }
)

# 2. Create and run the flow
task = Task(TextTaskInput("What is the latest update on GPT-5.2?"), search_agent)
flow = SequenceFlow([task])
result = flow.start()

print(result["task1"])

This mode searches through your own documents previously indexed via Intellicloud.

# Define the Intellicloud Search Agent
search_agent = Agent(
    agent_type=AgentTypes.SEARCH.value,
    provider="intellicloud",
    mission="search in my documentation",
    model_params={
        "one_key": "YOUR_INTELLICLOUD_ONE_KEY",
        "k": 3
    }
)

# Execution follows the same pattern as above

Notes​

  • Input: The Search Agent always expects a text-based query as input.
  • Output: By default, it returns a text summary of the results, making it easy to feed into a subsequent LLM task for analysis.