File size: 3,525 Bytes
1521ce5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 | from __future__ import annotations
import re
def check_answer_in_evidence(answer: str, evidence_sessions: list[dict]) -> bool:
evidence_text = ""
for sess in evidence_sessions:
for turn in sess.get("turns", []):
evidence_text += " " + turn["content"]
answer_lower = answer.lower().strip()
return answer_lower in evidence_text.lower()
def check_question_no_answer_leak(question: str, answer: str) -> bool:
"""Heuristic answer-leak guard.
Counts only "content" tokens — strips common stopwords AND time/measurement
units that are routinely shared between question and answer for two_hop /
temp_reasoning ("How many DAYS passed..." → "5 DAYS"). Without unit
stripping, ~all duration / count answers got flagged.
Threshold: leak_ratio >= 0.6 is rejected. We bumped from 0.5 → 0.6 so that
a single shared content token in a 2-token answer ("25 minutes") doesn't
auto-fail.
"""
stopwords = {
"the", "a", "an", "is", "was", "are", "were", "i", "my", "me", "do",
"what", "which", "how", "when", "where", "that", "this", "it",
"of", "to", "in", "on", "at", "for", "with", "from", "by", "and", "or",
}
unit_words = {
"day", "days", "week", "weeks", "month", "months", "year", "years",
"hour", "hours", "minute", "minutes", "second", "seconds",
"dollar", "dollars", "$", "cent", "cents",
"mile", "miles", "kilometer", "kilometers", "km", "meter", "meters",
"pound", "pounds", "kg", "kilogram", "kilograms",
"time", "times", "ago", "before", "after", "between",
}
skip = stopwords | unit_words
def tokens(text):
return {t.strip(".,!?;:'\"()") for t in text.lower().split()}
answer_tokens = tokens(answer) - skip
question_tokens = tokens(question) - skip
if not answer_tokens:
return True
overlap = answer_tokens & question_tokens
leak_ratio = len(overlap) / len(answer_tokens)
return leak_ratio < 0.6
def check_rewrite_no_answer_leak(rewrite_query: str, answer: str) -> bool:
if not rewrite_query or rewrite_query == "none":
return True
return check_question_no_answer_leak(rewrite_query, answer)
def check_entity_in_predictions(rewrite_query: str, predictions: list[str]) -> bool:
if not rewrite_query or rewrite_query == "none":
return True
pred_text = " ".join(predictions).lower()
rewrite_tokens = rewrite_query.lower().split()
stopwords = {"the", "a", "an", "is", "was", "are", "were", "i", "my", "me",
"what", "which", "how", "when", "where", "for", "to", "of", "in",
"and", "or", "on", "at", "by", "with", "from"}
meaningful_tokens = [t for t in rewrite_tokens if t not in stopwords and len(t) > 2]
if not meaningful_tokens:
return True
found = sum(1 for t in meaningful_tokens if t in pred_text)
return found / len(meaningful_tokens) >= 0.5
def check_no_user_specific_info(rewrite_query: str, original_query: str) -> bool:
if not rewrite_query or rewrite_query == "none":
return True
original_tokens = set(original_query.lower().split())
rewrite_tokens = set(rewrite_query.lower().split())
new_tokens = rewrite_tokens - original_tokens
stopwords = {"the", "a", "an", "is", "was", "are", "were", "i", "my", "me",
"what", "which", "how", "when", "where", "for", "to", "of", "in"}
meaningful_new = new_tokens - stopwords
return len(meaningful_new) <= 3
|