Skip to content

README.md

Latest commit

 

History

History
226 lines (169 loc) · 7.6 KB

README.md

File metadata and controls

226 lines (169 loc) · 7.6 KB
Raw

Letta Agent as Judge Example

This example demonstrates using a Letta agent as a judge for rubric-based grading.

What This Example Shows

  • Using a Letta agent as an LLM judge instead of direct API calls
  • Defining custom rubric criteria for evaluation
  • Validating judge agent tool schemas on initialization
  • Testing agent web-fetching capabilities with nuanced evaluation

How It Works

Instead of calling an LLM API directly (like the standard rubric grader), this approach:

  1. Loads a judge agent from a .af file
  2. Sends the formatted rubric prompt to the judge agent
  3. The judge agent evaluates the submission and calls a submit_grade tool with score and rationale
  4. Extracts the score from the tool call and uses it for grading

Key Components

Judge Agent (judge.af)

A Letta agent configured with:

  • A submit_grade(score: float, rationale: str) tool for submitting evaluations
  • Access to web fetching if needed for verification
  • System instructions for evaluation behavior

Rubric (```rubric.txt`)

Defines evaluation criteria:

  • Correctness: Does the response contain accurate information?
  • Format: Is the response properly formatted as requested?
  • Completeness: Does it fully address the question?

Dataset (dataset.csv)

Contains test cases where agents must fetch webpage content and answer questions.

Running This Example

Local Setup (Default)

Start your local Letta server:

letta server

Then run the evaluation:

cd examples/letta-agent-rubric-grader
letta-evals run suite.yaml

Letta Cloud Setup

Set these environment variables:

export LETTA_API_KEY=your-api-key
export LETTA_PROJECT_ID=your-project-id

Update base_url in suite.yaml:

target:
  base_url: https://api.letta.com/

Then run the evaluation as above.

Configuration Details

Suite Configuration

This example provides two configuration approaches:

Option 1: Default Letta Judge (Recommended for Most Use Cases)

File: default_judge_suite.yaml

The simplest configuration uses the built-in default judge agent with pre-fetched webpage content:

name: fetch-webpage-default-judge-test
dataset: dataset.csv
target:
  kind: letta_agent
  agent_file: test-fetch-webpage-simple-agent.af
graders:
  agent_judge:
    kind: letta_judge                      # Use letta_judge kind
    prompt_path: default_judge_rubric.txt  # Rubric with hardcoded webpage content
    extractor: last_assistant
gate:
  metric_key: agent_judge
  op: gte
  value: 0.7

How it works:

  • Uses the default Letta judge agent (no custom agent_file needed)
  • Rubric includes the hardcoded webpage content for grading
  • Judge reads the content directly from the rubric prompt
  • Simpler, faster, and more reliable (no live web requests)

When to use: Most evaluation scenarios where you know the expected content ahead of time.

Option 2: Custom Letta Judge with Live Web Search (Advanced)

File: suite.yaml

For advanced scenarios where the judge needs to dynamically verify information:

name: fetch-webpage-agent-judge-test
dataset: dataset.csv
target:
  kind: letta_agent
  agent_file: test-fetch-webpage-simple-agent.af
graders:
  agent_judge:
    kind: letta_judge
    agent_file: custom_web_search_judge.af  # Custom judge with web tools
    prompt_path: custom_judge_rubric.txt    # Rubric instructs live fetching
    judge_tool_name: submit_grade
    extractor: last_assistant
gate:
  metric_key: agent_judge
  op: gte
  value: 0.7

How it works:

  • Uses a custom judge agent with fetch_webpage tool capabilities
  • Rubric instructs the judge to fetch the webpage live during grading
  • Judge performs real-time web requests to verify agent answers
  • More dynamic but slower and depends on network availability

When to use: When evaluating against dynamic content, testing web-fetching capabilities, or when ground truth can't be pre-determined.

Key Differences:

Aspect Default Judge Custom Judge
Agent Built-in default Custom with web tools
Rubric Hardcoded content Instructions to fetch live
Speed Faster (no web requests) Slower (live fetching)
Reliability Higher (offline) Lower (network dependent)
Use Case Static evaluation Dynamic verification
Config Complexity Minimal (2 required fields) Higher (4+ fields)

Key Configuration Options:

  • kind: Must be letta_judge for agent-based judges
  • agent_file: (Optional) Path to custom .af judge agent. If omitted, uses default judge
  • prompt_path: Path to file containing rubric text (can also use prompt for inline rubric)
  • judge_tool_name: (Optional) Name of the tool the judge calls to submit scores. Only allowed with custom agent_file
  • extractor: How to extract the submission from agent trajectory

Judge Agent Requirements & Gotchas

Recommendation: Use the Default Judge

We highly recommend using the default Letta judge (Option 1) for most use cases. Configuring a custom judge agent is complex and error-prone, with several potential footguns:

  • Tool schema must exactly match expected parameters
  • Tool name must be correctly specified
  • Agent must not have conflicting tools that confuse evaluation
  • Additional complexity in debugging when things go wrong

Only create a custom judge if you have a specific need, such as:

  • Judge needs to fetch live web content for verification
  • Judge requires access to custom tools (databases, APIs, etc.)
  • Special evaluation logic that can't be expressed in the rubric alone

If you do need a custom judge, use this checklist:

Checklist: Will Your Judge Agent Work?

Use this checklist to verify your judge agent is properly configured:

  • Tool exists: Agent has a tool with the name specified in judge_tool_name (default: submit_grade)
  • Tool parameters: The tool has BOTH score: float and rationale: str parameters
  • Tool is callable: The tool is not disabled or requires-approval-only
  • Agent system prompt: Agent understands it's an evaluator (optional but recommended)
  • No conflicting tools: Agent doesn't have other tools that might confuse it into answering questions instead of judging

Required Tool Signature

The judge agent must have a tool with this exact parameter signature:

def submit_grade(score: float, rationale: str) -> dict:
    """
    Submit an evaluation grade for an agent's response.

    Args:
        score: A float between 0.0 (complete failure) and 1.0 (perfect)
        rationale: Explanation of why this score was given

    Returns:
        dict: Confirmation of grade submission
    """
    return {
        "status": "success",
        "grade": {"score": score, "rationale": rationale}
    }

Important: The parameter names must be exactly score and rationale. The framework validates this on initialization.

Validation on Initialization

The framework validates your judge agent before running any evaluations. If validation fails, you'll get a clear error message:

ValueError: Judge tool 'submit_grade' not found in agent file judge.af.
Available tools: ['fetch_webpage', 'search_documents']

This fail-fast approach saves time by catching configuration errors immediately.

Dataset Format

input
"Read `https://www.york.ac.uk/teaching/cws/wws/webpage1.html`. What program is mentioned for writing HTML code? Respond with the program name ONLY in brackets, e.g. {Word}."

Note: Ground truth is optional for agent judges since the judge agent can verify answers independently (e.g., by fetching the webpage itself).