CrewAI Integration

GroundlensTool enables CrewAI agents to verify their own outputs for hallucination risk before presenting results to users or other agents.

Installation

pip install "groundlens[crewai]"

Quick Start

from groundlens.integrations.crewai import GroundlensTool

tool = GroundlensTool()

# Verify a response
result = tool._run(
    question="What is the recommended dosage for ibuprofen?",
    response="The recommended dosage is 200-400mg every 4-6 hours.",
    context="Ibuprofen: 200-400mg PO every 4-6 hours. Max 1200mg/day OTC.",
)
print(result)

Output:

Groundlens Verification Result
----------------------------
Method: SGI
Score: 1.234 (normalized: 0.614)
Status: PASS
Explanation: SGI=1.234 -- strong context engagement (pass)

Configuration

tool = GroundlensTool(
    name="groundlens_verify",           # Tool name visible to the agent
    description="Verify a response...",  # Custom description
    groundlens_model="all-MiniLM-L6-v2",  # Embedding model
)

Parameter	Default	Description
name	"groundlens_verify"	Tool name for agent tool selection
description	(built-in)	Description shown to the agent
groundlens_model	"all-MiniLM-L6-v2"	Sentence-transformer for scoring

Using in a CrewAI Agent

from crewai import Agent, Task, Crew
from groundlens.integrations.crewai import GroundlensTool

# Create the verification tool
verify_tool = GroundlensTool()

# Create an agent that uses groundlens for self-verification
researcher = Agent(
    role="Research Analyst",
    goal="Provide accurate, verified research summaries",
    tools=[verify_tool],
    backstory="You are a meticulous researcher who always verifies your findings.",
)

# The agent can call groundlens_verify to check its own outputs
task = Task(
    description="Research the effects of caffeine on sleep quality. Verify your findings.",
    agent=researcher,
    expected_output="A verified summary of caffeine's effects on sleep.",
)

crew = Crew(agents=[researcher], tasks=[task])
result = crew.kickoff()

Tool Input/Output

Input

The _run method accepts:

Parameter	Required	Description
question	Yes	The original question or prompt
response	Yes	The response to verify
context	No	Source context (enables SGI when provided)

Output

Returns a formatted string containing:

• Method: SGI or DGI
• Score: Raw and normalized values
• Status: PASS or FLAGGED
• Explanation: Human-readable interpretation
• Recommendation (if flagged): Suggestion to revise with verified sources

Agent Self-Verification Pattern

The most powerful use of GroundlensTool is agent self-verification: the agent generates a response, verifies it with groundlens, and revises if flagged.

researcher = Agent(
    role="Medical Information Specialist",
    goal="Provide accurate medical information, always verified",
    tools=[GroundlensTool()],
    backstory=(
        "You are a medical information specialist. Before presenting any answer, "
        "you MUST use the groundlens_verify tool to check your response. If flagged, "
        "revise your answer and verify again."
    ),
)