Build latest book artifacts / build (push) Canceled after 0s
dependency resolution / resolve (3.11) (push) Canceled after 0s
dependency resolution / resolve (3.13) (push) Canceled after 0s
deploy-pages / build (push) Canceled after 0s
deploy-pages / deploy (push) Canceled after 0s
i18n consistency check / check (push) Canceled after 0s
provider adoption tests / test (chapter2/context-compression) (push) Canceled after 0s
provider adoption tests / test (chapter2/prompt-injection) (push) Canceled after 0s
provider adoption tests / test (chapter2/system-hint) (push) Canceled after 0s
provider adoption tests / test (chapter3/log-sanitization) (push) Canceled after 0s
web-search-agent tests / test (push) Canceled after 0s
web-search-agent tests / agentbook (push) Canceled after 0s
125 lines
4.5 KiB
Python
125 lines
4.5 KiB
Python
"""Offline, deterministic metrics for the User Memory Evaluation Framework.
|
|
|
|
The LLM-as-judge evaluator in ``evaluator.py`` requires an API key and network
|
|
access. This module provides a complementary metric that runs fully offline on
|
|
canned data, so the benchmark can produce a scored comparison across memory
|
|
systems without calling any LLM.
|
|
|
|
The ``KeywordRecallEvaluator`` implements *key-fact recall*: for each test case a
|
|
set of gold facts (account numbers, confirmation codes, entity names, dates that
|
|
were actually stated in the conversation histories) is checked for presence in
|
|
the agent's response via normalized substring matching. The reward equals the
|
|
fraction of gold facts recalled, i.e. a classic answer-contains-gold recall
|
|
metric. It shares the ``EvaluationResult`` output shape with ``LLMEvaluator`` so
|
|
both metrics are interchangeable in the reporting/comparison code.
|
|
"""
|
|
|
|
import json
|
|
import re
|
|
from pathlib import Path
|
|
from typing import Dict, List, Optional, Union
|
|
|
|
from models import TestCase, EvaluationResult
|
|
|
|
|
|
# A gold fact is either a single required string, or a list of acceptable
|
|
# variants (any one of which counts as a match, e.g. ["Feb 18", "February 18"]).
|
|
GoldFact = Union[str, List[str]]
|
|
|
|
|
|
def _normalize(text: str) -> str:
|
|
"""Lowercase and collapse whitespace for tolerant substring matching."""
|
|
return re.sub(r"\s+", " ", (text or "").lower()).strip()
|
|
|
|
|
|
def _fact_matched(fact: GoldFact, normalized_response: str) -> bool:
|
|
"""Return True if the (possibly multi-variant) fact appears in the response."""
|
|
variants = fact if isinstance(fact, list) else [fact]
|
|
return any(_normalize(v) in normalized_response for v in variants)
|
|
|
|
|
|
def _fact_label(fact: GoldFact) -> str:
|
|
"""Human-readable label for a gold fact (first variant for any-of facts)."""
|
|
if isinstance(fact, list):
|
|
return fact[0] + (" (…)" if len(fact) > 1 else "")
|
|
return fact
|
|
|
|
|
|
def load_gold_facts(path: Union[str, Path]) -> Dict[str, List[GoldFact]]:
|
|
"""Load gold-fact annotations from a JSON file.
|
|
|
|
The file maps ``test_id`` to an object with a ``required_facts`` list. Each
|
|
entry in ``required_facts`` is a string, or a list of acceptable variants.
|
|
"""
|
|
with open(path, "r", encoding="utf-8") as f:
|
|
data = json.load(f)
|
|
|
|
gold: Dict[str, List[GoldFact]] = {}
|
|
for test_id, spec in data.items():
|
|
if isinstance(spec, dict):
|
|
gold[test_id] = spec.get("required_facts", [])
|
|
elif isinstance(spec, list):
|
|
gold[test_id] = spec
|
|
return gold
|
|
|
|
|
|
class KeywordRecallEvaluator:
|
|
"""Offline key-fact recall metric (no LLM / API required)."""
|
|
|
|
name = "keyword-recall"
|
|
|
|
def __init__(self, gold_facts: Dict[str, List[GoldFact]]):
|
|
"""
|
|
Args:
|
|
gold_facts: Mapping of test_id -> list of gold facts to recall.
|
|
"""
|
|
self.gold_facts = gold_facts
|
|
|
|
def has_gold(self, test_id: str) -> bool:
|
|
"""Whether gold facts are available for the given test case."""
|
|
return bool(self.gold_facts.get(test_id))
|
|
|
|
def evaluate(
|
|
self,
|
|
test_case: TestCase,
|
|
agent_response: str,
|
|
extracted_memory: Optional[str] = None,
|
|
) -> EvaluationResult:
|
|
"""Score a response by the fraction of gold facts it recalls.
|
|
|
|
The optional ``extracted_memory`` is concatenated with the response so a
|
|
fact stated in the agent's memory dump also counts as recalled.
|
|
"""
|
|
facts = self.gold_facts.get(test_case.test_id, [])
|
|
haystack = _normalize(f"{agent_response}\n{extracted_memory or ''}")
|
|
|
|
if not facts:
|
|
return EvaluationResult(
|
|
test_id=test_case.test_id,
|
|
reward=0.0,
|
|
passed=None,
|
|
reasoning="No gold facts defined for this test case; skipped by keyword-recall metric.",
|
|
required_info_found={},
|
|
)
|
|
|
|
info_found: Dict[str, float] = {}
|
|
matched = 0
|
|
for fact in facts:
|
|
hit = _fact_matched(fact, haystack)
|
|
info_found[_fact_label(fact)] = 1.0 if hit else 0.0
|
|
matched += int(hit)
|
|
|
|
recall = matched / len(facts)
|
|
missing = [label for label, score in info_found.items() if score == 0.0]
|
|
reasoning = f"Recalled {matched}/{len(facts)} gold facts."
|
|
if missing:
|
|
reasoning += " Missing: " + ", ".join(missing) + "."
|
|
|
|
return EvaluationResult(
|
|
test_id=test_case.test_id,
|
|
reward=recall,
|
|
passed=recall >= 0.8,
|
|
reasoning=reasoning,
|
|
required_info_found=info_found,
|
|
)
|