Spaces:
Running
Running
File size: 14,490 Bytes
8299003 | 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 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 | """
verifier.py
===========
Phase 7 β Answer Grounding & Hallucination Detection
Verifies that every sentence in an LLM-generated answer is supported by the
retrieved context chunks. Uses a Natural Language Inference (NLI) cross-encoder
to classify each (context, claim) pair as:
SUPPORTED β context entails the claim (entailment score β₯ threshold)
UNVERIFIED β context neither supports nor contradicts (neutral)
CONTRADICTED β context explicitly contradicts the claim
NLI Model
---------
cross-encoder/nli-deberta-v3-small
- ~180 MB, runs on CPU
- Label order: {0: contradiction, 1: entailment, 2: neutral}
- Input: (premise=context_chunk, hypothesis=answer_sentence)
Two-step verification process
------------------------------
1. Sentence splitting
Split the LLM answer into individual claims using NLTK sentence tokenizer.
2. Claim-context entailment
For each sentence, pair it against every retrieved context chunk.
The chunk with the highest entailment score is the best support.
Classify the sentence based on that best score.
3. Citation verification (optional)
Extract [1], [2], ... references from the answer.
Check that the cited chunk actually supports the citing sentence.
Grounding score
---------------
grounding_score = supported_sentences / total_sentences
Range: 0.0 (fully hallucinated) β 1.0 (fully grounded)
Typical acceptable threshold: β₯ 0.7
Usage
-----
from src.verifier import AnswerVerifier
verifier = AnswerVerifier()
result = verifier.verify(answer=answer_text, chunks=retrieved_chunks)
print(f"Grounding score: {result['grounding_score']:.0%}")
for r in result["sentence_results"]:
verdict = r["verdict"]
sent = r["sentence"][:80]
print(f" [{verdict:12s}] {sent}")
# Citation check
cites = verifier.check_citations(answer=answer_text, chunks=retrieved_chunks)
for c in cites:
print(f" {c['citation']} {c['status']} entail={c.get('entail_score','n/a')}")
"""
import logging
import re
from pathlib import Path
# ββ Logging ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
logging.basicConfig(
level = logging.INFO,
format = "%(asctime)s %(levelname)-8s %(message)s",
)
log = logging.getLogger(__name__)
# ββ Constants ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
NLI_MODEL = "cross-encoder/nli-deberta-v3-small"
ENTAIL_IDX = 1 # confirmed: {0: contradiction, 1: entailment, 2: neutral}
CONTRA_IDX = 0
NEUTRAL_IDX = 2
ENTAIL_THRESHOLD = 0.50 # sentence classified as SUPPORTED if entail β₯ this
CONTRA_THRESHOLD = 0.40 # sentence classified as CONTRADICTED if contra β₯ this
MIN_SENTENCE_LEN = 20 # shorter fragments are skipped (headers, bullet markers)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# ANSWER VERIFIER
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class AnswerVerifier:
"""
Verifies LLM answer grounding against retrieved context chunks using NLI.
Attributes
----------
entail_threshold : float
Minimum entailment probability for a sentence to be SUPPORTED.
contra_threshold : float
Minimum contradiction probability to flag a sentence as CONTRADICTED.
"""
def __init__(
self,
nli_model : str = NLI_MODEL,
entail_threshold : float = ENTAIL_THRESHOLD,
contra_threshold : float = CONTRA_THRESHOLD,
):
from sentence_transformers import CrossEncoder
log.info(f"Loading NLI model: {nli_model}")
self._nli = CrossEncoder(nli_model, max_length=512)
self.entail_threshold = entail_threshold
self.contra_threshold = contra_threshold
log.info("NLI model ready.")
# ββ Sentence splitting ββββββββββββββββββββββββββββββββββββββββββββββββββ
def split_sentences(self, text: str) -> list[str]:
"""
Split answer text into individual sentences.
Uses NLTK sent_tokenize for reliable sentence boundary detection.
Filters out very short fragments (bullet markers, standalone numbers).
"""
try:
import nltk
try:
sentences = nltk.sent_tokenize(text)
except LookupError:
nltk.download("punkt_tab", quiet=True)
sentences = nltk.sent_tokenize(text)
except ImportError:
# Simple regex fallback if NLTK unavailable
sentences = re.split(r"(?<=[.!?])\s+(?=[A-Z\[\(])", text.strip())
return [s.strip() for s in sentences if len(s.strip()) >= MIN_SENTENCE_LEN]
# ββ Core verification βββββββββββββββββββββββββββββββββββββββββββββββββββ
def verify(
self,
answer : str,
chunks : list[dict],
verbose : bool = False,
) -> dict:
"""
Verify every sentence in `answer` against the retrieved `chunks`.
Args:
answer : LLM-generated answer string
chunks : list of chunk dicts with keys: id, text, metadata, score
verbose : if True, log each sentence verdict
Returns:
{
"grounding_score" : float (0.0β1.0),
"total_sentences" : int,
"supported" : int,
"unverified" : int,
"contradicted" : int,
"sentence_results" : list[dict],
}
"""
# ββ Check for prescribed "not found" phrase ββββββββββββββββββββββββ
NOT_FOUND_PHRASES = [
"not contain enough information",
"not available in the provided",
"cannot find",
"no information",
]
if any(p in answer.lower() for p in NOT_FOUND_PHRASES):
log.info("Answer contains explicit 'not found' phrase β trivially grounded.")
return {
"grounding_score" : 1.0,
"total_sentences" : 0,
"supported" : 0,
"unverified" : 0,
"contradicted" : 0,
"sentence_results" : [],
"note" : "LLM correctly reported no relevant context found.",
}
sentences = self.split_sentences(answer)
context_texts = [c["text"] for c in chunks]
if not sentences:
return {
"grounding_score": 0.0, "total_sentences": 0,
"supported": 0, "unverified": 0, "contradicted": 0,
"sentence_results": [],
}
sentence_results = []
for sent in sentences:
# Pair each sentence against ALL context chunks as premises
pairs = [(ctx, sent) for ctx in context_texts]
scores = self._nli.predict(pairs, apply_softmax=True)
# Find the chunk with the highest entailment score for this sentence
best_idx = max(range(len(scores)), key=lambda i: float(scores[i][ENTAIL_IDX]))
best_scores = scores[best_idx]
entail_prob = float(best_scores[ENTAIL_IDX])
contra_prob = float(best_scores[CONTRA_IDX])
neutral_prob = float(best_scores[NEUTRAL_IDX])
# Classify
if entail_prob >= self.entail_threshold:
verdict = "SUPPORTED"
elif contra_prob >= self.contra_threshold:
verdict = "CONTRADICTED"
else:
verdict = "UNVERIFIED"
result = {
"sentence" : sent,
"verdict" : verdict,
"entail_prob" : round(entail_prob, 3),
"contra_prob" : round(contra_prob, 3),
"neutral_prob" : round(neutral_prob, 3),
"best_chunk_id" : chunks[best_idx]["id"],
"best_chunk_text" : context_texts[best_idx][:120],
}
sentence_results.append(result)
if verbose:
log.info(
f" [{verdict:12s}] entail={entail_prob:.2f} "
f"contra={contra_prob:.2f} | {sent[:70]!r}"
)
supported = sum(1 for r in sentence_results if r["verdict"] == "SUPPORTED")
unverified = sum(1 for r in sentence_results if r["verdict"] == "UNVERIFIED")
contradicted = sum(1 for r in sentence_results if r["verdict"] == "CONTRADICTED")
grounding = supported / len(sentence_results)
return {
"grounding_score" : round(grounding, 3),
"total_sentences" : len(sentence_results),
"supported" : supported,
"unverified" : unverified,
"contradicted" : contradicted,
"sentence_results" : sentence_results,
}
# ββ Citation verification βββββββββββββββββββββββββββββββββββββββββββββββ
def check_citations(
self,
answer : str,
chunks : list[dict],
) -> list[dict]:
"""
Verify that [1], [2], ... citations in the answer refer to the right chunk.
For each citation, checks whether the cited chunk actually entails the
sentence containing the citation.
Args:
answer : LLM-generated answer with inline citations like [1], [2]
chunks : list of chunk dicts (same order as context was assembled)
Returns:
list of dicts:
citation : "[1]"
sentence : the sentence containing the citation (first 120 chars)
chunk_id : ID of the cited chunk
entail_score : how well the chunk supports the sentence
status : "CORRECT" | "QUESTIONABLE" | "OUT_OF_RANGE" | "NO_CITATIONS"
"""
sentences = self.split_sentences(answer)
results = []
for sent in sentences:
cite_nums = re.findall(r"\[(\d+)\]", sent)
for num_str in cite_nums:
idx = int(num_str) - 1 # citations are 1-indexed
if idx < 0 or idx >= len(chunks):
results.append({
"citation" : f"[{num_str}]",
"sentence" : sent[:120],
"chunk_id" : None,
"entail_score": None,
"status" : "OUT_OF_RANGE",
})
continue
chunk_text = chunks[idx]["text"]
scores = self._nli.predict(
[(chunk_text, sent)], apply_softmax=True
)
entail_prob = round(float(scores[0][ENTAIL_IDX]), 3)
results.append({
"citation" : f"[{num_str}]",
"sentence" : sent[:120],
"chunk_id" : chunks[idx]["id"],
"entail_score": entail_prob,
"status" : "CORRECT" if entail_prob >= 0.35 else "QUESTIONABLE",
})
if not results:
return [{"status": "NO_CITATIONS", "note": "Answer contains no [n] citations."}]
return results
# ββ Summary report ββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def report(self, verification_result: dict, citation_result: list[dict] = None) -> str:
"""
Format a human-readable grounding report.
"""
r = verification_result
score = r["grounding_score"]
total = r["total_sentences"]
if score >= 0.85:
grade = "PASS β well-grounded"
elif score >= 0.60:
grade = "REVIEW β partially grounded"
else:
grade = "FAIL β high hallucination risk"
lines = [
"=" * 65,
f" Grounding Report",
"=" * 65,
f" Score : {score:.0%} ({r['supported']}/{total} sentences supported)",
f" Verdict : {grade}",
f" Breakdown: {r['supported']} supported | "
f"{r['unverified']} unverified | {r['contradicted']} contradicted",
"-" * 65,
]
for i, s in enumerate(r["sentence_results"], 1):
icon = {"SUPPORTED": "β", "UNVERIFIED": "?", "CONTRADICTED": "β"}.get(
s["verdict"], " "
)
lines.append(
f" [{i}] {icon} [{s['verdict']:12s}] "
f"e={s['entail_prob']:.2f} c={s['contra_prob']:.2f}"
)
lines.append(f" {s['sentence'][:90]}")
lines.append(f" β best match: {s['best_chunk_id']}")
if citation_result:
lines += ["", "-" * 65, " Citation Check"]
for c in citation_result:
if c.get("status") == "NO_CITATIONS":
lines.append(" No inline citations found in answer.")
else:
score_str = f"entail={c['entail_score']:.2f}" if c["entail_score"] else "n/a"
lines.append(
f" {c['citation']} [{c['status']:12s}] {score_str} "
f"chunk={c['chunk_id']}"
)
lines.append("=" * 65)
return "\n".join(lines)
|