Richard T
111 lines
3.2 KiB
Python
111 lines
3.2 KiB
Python
"""
|
|
LLM-based judge for semantic evaluation of test results.
|
|
|
|
Used by tests that need to evaluate semantic properties like
|
|
"no hallucination" or "preserves meaning" that can't be checked
|
|
with simple assertions.
|
|
|
|
Usage in tests:
|
|
from framework.testing.llm_judge import LLMJudge
|
|
|
|
judge = LLMJudge()
|
|
result = judge.evaluate(
|
|
constraint="no-hallucination",
|
|
source_document="The original text...",
|
|
summary="The summary to evaluate...",
|
|
criteria="Summary must only contain facts from the source"
|
|
)
|
|
assert result["passes"], result["explanation"]
|
|
"""
|
|
|
|
import json
|
|
from typing import Any
|
|
|
|
|
|
class LLMJudge:
|
|
"""
|
|
LLM-based judge for semantic evaluation of test results.
|
|
|
|
Uses Claude to evaluate whether outputs meet semantic constraints
|
|
that can't be verified with simple assertions.
|
|
"""
|
|
|
|
def __init__(self):
|
|
"""Initialize the LLM judge."""
|
|
self._client = None
|
|
|
|
def _get_client(self):
|
|
"""Lazy-load the Anthropic client."""
|
|
if self._client is None:
|
|
try:
|
|
import anthropic
|
|
|
|
self._client = anthropic.Anthropic()
|
|
except ImportError as e:
|
|
raise RuntimeError("anthropic package required for LLM judge") from e
|
|
return self._client
|
|
|
|
def evaluate(
|
|
self,
|
|
constraint: str,
|
|
source_document: str,
|
|
summary: str,
|
|
criteria: str,
|
|
) -> dict[str, Any]:
|
|
"""
|
|
Evaluate whether a summary meets a constraint.
|
|
|
|
Args:
|
|
constraint: The constraint being tested (e.g., "no-hallucination")
|
|
source_document: The original document
|
|
summary: The generated summary to evaluate
|
|
criteria: Human-readable criteria for evaluation
|
|
|
|
Returns:
|
|
Dict with 'passes' (bool) and 'explanation' (str)
|
|
"""
|
|
client = self._get_client()
|
|
|
|
prompt = f"""You are evaluating whether a summary meets a specific constraint.
|
|
|
|
CONSTRAINT: {constraint}
|
|
CRITERIA: {criteria}
|
|
|
|
SOURCE DOCUMENT:
|
|
{source_document}
|
|
|
|
SUMMARY TO EVALUATE:
|
|
{summary}
|
|
|
|
Evaluate whether the summary meets the constraint. Be strict but fair.
|
|
|
|
Respond with JSON in this exact format:
|
|
{{"passes": true/false, "explanation": "brief explanation of your judgment"}}
|
|
|
|
Only output the JSON, nothing else."""
|
|
|
|
try:
|
|
response = client.messages.create(
|
|
model="claude-haiku-4-5-20251001",
|
|
max_tokens=500,
|
|
messages=[{"role": "user", "content": prompt}],
|
|
)
|
|
|
|
# Parse the response
|
|
text = response.content[0].text.strip()
|
|
# Handle potential markdown code blocks
|
|
if text.startswith("```"):
|
|
text = text.split("```")[1]
|
|
if text.startswith("json"):
|
|
text = text[4:]
|
|
text = text.strip()
|
|
|
|
result = json.loads(text)
|
|
return {
|
|
"passes": bool(result.get("passes", False)),
|
|
"explanation": result.get("explanation", "No explanation provided"),
|
|
}
|
|
except Exception as e:
|
|
# On error, fail the test with explanation
|
|
return {"passes": False, "explanation": f"LLM judge error: {e}"}
|