Reconstruct optimized verifier stack

Author: Reyaansh Sinha

Halgorithem/__init__.py

@@ -1,3 +1,4 @@
 from .core import Halgorithm
+from .main import HalgorithemVerifier, verify, verify_urls
 
-__all__ = ["Halgorithm"]
+__all__ = ["Halgorithm", "HalgorithemVerifier", "verify", "verify_urls"]

Halgorithem/checks/__init__.py

@@ -0,0 +1 @@
+"""Higher-level verification checks for Halgorithem."""

Halgorithem/checks/atomic.py

@@ -0,0 +1,91 @@
+import re
+from functools import lru_cache
+
+from ..claim_extraction import split_atomic_claims
+from ..models import AtomicCheck, AtomicClaim
+
+
+@lru_cache(maxsize=8192)
+def _tokens(text):
+    return frozenset(t.lower() for t in re.findall(r"\b[a-zA-Z][a-zA-Z'-]+\b", text or "") if len(t) > 2)
+
+
+def _overlap(left, right):
+    if not left:
+        return 0.0
+    return len(left & right) / len(left)
+
+
+def _priority(verdict):
+    return {"CONTRADICT": 3, "ENTAIL": 2, "NEUTRAL": 1}.get(verdict, 0)
+
+
+def _score_claim(verdict, confidence):
+    if verdict == "ENTAIL":
+        return confidence
+    if verdict == "CONTRADICT":
+        return -confidence
+    return 0.5 * confidence
+
+
+def prepare_document_claims(document):
+    doc_claims = []
+    for sentence in document:
+        for claim in split_atomic_claims(sentence.resolved_text) or [sentence.resolved_text]:
+            doc_claims.append((claim, _tokens(claim)))
+    return doc_claims
+
+
+def atomic_claim_nli(processed_sentence, document, nli_model=None, doc_claims=None):
+    ai_text = getattr(processed_sentence, "resolved_text", processed_sentence)
+    ai_claims = split_atomic_claims(ai_text) or [ai_text]
+    doc_claims = doc_claims if doc_claims is not None else prepare_document_claims(document)
+
+    if not doc_claims:
+        return AtomicCheck(status="no_document_claims")
+
+    matched = []
+    for claim in ai_claims:
+        claim_tokens = _tokens(claim)
+        best_claim, best_tokens = max(doc_claims, key=lambda pair: _overlap(claim_tokens, pair[1]))
+        matched.append((claim, best_claim, best_tokens))
+
+    if nli_model is not None:
+        nli_results = nli_model.predict_batch(
+            [best_claim for _, best_claim, _ in matched],
+            [claim for claim, _, _ in matched],
+        )
+    else:
+        nli_results = [None] * len(matched)
+
+    results = []
+    for (claim, best_claim, best_tokens), nli in zip(matched, nli_results):
+        claim_tokens = _tokens(claim)
+        if not claim_tokens or not best_tokens:
+            results.append(AtomicClaim(claim=claim, verdict="NEUTRAL", confidence=0.0, evidence=""))
+            continue
+        if nli is None:
+            overlap = _overlap(claim_tokens, best_tokens)
+            verdict = "ENTAIL" if overlap >= 0.70 else "NEUTRAL"
+            confidence = overlap if verdict == "ENTAIL" else 0.50
+        elif nli.label == "CONTRADICTION":
+            verdict = "CONTRADICT"
+            confidence = nli.score
+        elif nli.label == "ENTAILMENT":
+            verdict = "ENTAIL"
+            confidence = nli.score
+        else:
+            verdict = "NEUTRAL"
+            confidence = nli.score
+        results.append(AtomicClaim(claim=claim, verdict=verdict, confidence=confidence, evidence=best_claim))
+
+    entail = sum(1 for c in results if c.verdict == "ENTAIL")
+    contradict = sum(1 for c in results if c.verdict == "CONTRADICT")
+    total = len(results)
+    if total:
+        weighted_sum = sum(_score_claim(claim.verdict, claim.confidence) for claim in results)
+        score = weighted_sum / total
+    else:
+        score = None
+    results.sort(key=lambda c: (_priority(c.verdict), c.confidence), reverse=True)
+    return AtomicCheck(claims=results, score=score, status="ok")

Halgorithem/checks/nli.py

@@ -0,0 +1,101 @@
+import re
+from functools import lru_cache
+
+from ..contradiction import find_contradiction
+from ..models import NLICheck
+from ..text_processing import extract_numbers, has_negation_mismatch, lemmatize_tokens
+
+
+@lru_cache(maxsize=8192)
+def _tokens(text):
+    return frozenset(t.lower() for t in re.findall(r"\b[a-zA-Z][a-zA-Z'-]+\b", text or "") if len(t) > 2)
+
+
+@lru_cache(maxsize=8192)
+def _content_lemmas(text):
+    return frozenset(lemma for lemma in lemmatize_tokens(text) if len(lemma) > 2)
+
+
+class NLIModel:
+    model_quality = 0.75
+
+    def predict(self, premise, hypothesis):
+        return self.predict_batch([premise], [hypothesis])[0]
+
+    def predict_batch(self, premises, hypotheses):
+        return [rule_nli(premise, hypothesis) for premise, hypothesis in zip(premises, hypotheses)]
+
+
+def rule_nli(premise, hypothesis):
+    chunk = {"text": premise or "", "numbers": extract_numbers(premise)}
+    issue = find_contradiction(
+        claim=hypothesis,
+        chunk=chunk,
+        extract_numbers=extract_numbers,
+        has_negation_mismatch=has_negation_mismatch,
+        score=1.0,
+        threshold=0.0,
+    )
+    if issue:
+        return NLICheck("CONTRADICTION", 0.84, issue.get("reason", "Contradiction"), model_quality=0.75)
+
+    premise_tokens = _tokens(premise)
+    hypothesis_tokens = _tokens(hypothesis)
+    premise_numbers = set(extract_numbers(premise))
+    hypothesis_numbers = set(extract_numbers(hypothesis))
+    if hypothesis_numbers and not premise_numbers:
+        return NLICheck("NEUTRAL", 0.35, "Missing number evidence", model_quality=0.75)
+    if hypothesis_numbers and premise_numbers and not hypothesis_numbers.issubset(premise_numbers):
+        return NLICheck("NEUTRAL", 0.42, "Number evidence differs", model_quality=0.75)
+    if hypothesis_tokens:
+        token_overlap = len(premise_tokens & hypothesis_tokens) / len(hypothesis_tokens)
+        premise_lemmas = _content_lemmas(premise)
+        hypothesis_lemmas = _content_lemmas(hypothesis)
+        lemma_overlap = len(premise_lemmas & hypothesis_lemmas) / len(hypothesis_lemmas) if hypothesis_lemmas else 0.0
+        overlap = max(token_overlap, lemma_overlap)
+        if overlap >= 0.70:
+            return NLICheck("ENTAILMENT", min(0.60 + overlap * 0.30, 0.92), model_quality=0.75)
+    return NLICheck("NEUTRAL", 0.50, model_quality=0.75)
+
+
+def sentence_nli(processed_sentence, document=None, nli_model=None, hits=None):
+    model = nli_model or NLIModel()
+    claim = getattr(processed_sentence, "resolved_text", processed_sentence)
+    claim = claim if isinstance(claim, str) else str(claim)
+    relevant_hits = hits or []
+    if not relevant_hits and document is not None:
+        relevant_hits = [{"sentence": s.resolved_text, "score": 1.0} for s in document[:1]]
+
+    best_hit_score = max(
+        (
+            (hit.get("score", 0.0) if isinstance(hit, dict) else getattr(hit, "score", 0.0))
+            for hit in relevant_hits
+        ),
+        default=0.0,
+    )
+    min_hit_score = max(0.30, best_hit_score * 0.80)
+    premises = []
+    hypotheses = []
+    hit_scores = []
+    for hit in relevant_hits[:5]:
+        premise = hit.get("sentence") if isinstance(hit, dict) else getattr(hit, "sentence", str(hit))
+        hit_score = hit.get("score", 0.0) if isinstance(hit, dict) else getattr(hit, "score", 0.0)
+        if hit_score < min_hit_score:
+            continue
+        premises.append(premise)
+        hypotheses.append(claim)
+        hit_scores.append(hit_score)
+    if not premises:
+        return NLICheck("NEUTRAL", 0.50, model_quality=getattr(model, "model_quality", 1.0))
+
+    results = model.predict_batch(premises, hypotheses)
+    entailments = [r for r in results if r.label == "ENTAILMENT"]
+    strong_entailment = max(entailments, key=lambda r: r.score) if entailments else None
+    contradictions = [r for r in results if r.label == "CONTRADICTION"]
+    if strong_entailment and strong_entailment.score >= 0.82:
+        return strong_entailment
+    if contradictions:
+        return max(contradictions, key=lambda r: r.score)
+    if entailments:
+        return max(entailments, key=lambda r: r.score)
+    return max(results, key=lambda r: r.score)

Halgorithem/checks/similarity.py

@@ -0,0 +1,54 @@
+import heapq
+
+from ..models import SimilarityCheck, SimilarityHit
+def _token_overlap(left_tokens, right_tokens):
+    left_tokens = {token for token in left_tokens if len(token) > 2}
+    if not left_tokens or not right_tokens:
+        return 0.0
+    return len(left_tokens & set(right_tokens)) / len(left_tokens)
+
+
+def _similarities(embedder, left, rights):
+    if hasattr(embedder, "similarity_many"):
+        try:
+            return embedder.similarity_many(left, rights)
+        except Exception:
+            pass
+    return [float(embedder.similarity(left, right)) for right in rights]
+
+
+def similarity_search(processed_sentence, document, embedder, top_k=5):
+    query_embedding = processed_sentence.embedding
+    if query_embedding is None:
+        query_embedding = embedder.encode(processed_sentence.resolved_text, convert_to_tensor=True)
+
+    embeddings = []
+    for doc_sentence in document:
+        embedding = doc_sentence.embedding
+        if embedding is None:
+            embedding = embedder.encode(doc_sentence.resolved_text, convert_to_tensor=True)
+        embeddings.append(embedding)
+
+    raw_scores = _similarities(embedder, query_embedding, embeddings)
+    hits = []
+    for doc_sentence, raw_score in zip(document, raw_scores):
+        overlap = _token_overlap(processed_sentence.tokens, doc_sentence.tokens)
+        lemma_overlap = _token_overlap(processed_sentence.lemmas, doc_sentence.lemmas)
+        number_bonus = 0.05 if processed_sentence.numbers and processed_sentence.numbers.issubset(doc_sentence.numbers) else 0.0
+        overlap = max(overlap, lemma_overlap)
+        score = min(raw_score + 0.12 * overlap, 1.0)
+        score = min(score + number_bonus, 1.0)
+        hits.append(
+            SimilarityHit(
+                sentence=doc_sentence.context_text or doc_sentence.resolved_text,
+                score=score,
+                source=doc_sentence.source,
+                sentence_id=doc_sentence.sentence_id,
+                source_quality=doc_sentence.source_quality,
+            )
+        )
+
+    selected = heapq.nlargest(top_k, hits, key=lambda hit: hit.score)
+    source_quality = max((hit.source_quality for hit in selected), default=0.55)
+    score = selected[0].score if selected else 0.0
+    return SimilarityCheck(score=score, hits=selected, source_quality=source_quality)

Halgorithem/checks/units.py

@@ -0,0 +1,38 @@
+UNIT_ALIASES = {
+    "g": "gram",
+    "gram": "gram",
+    "grams": "gram",
+    "kg": "kilogram",
+    "kilogram": "kilogram",
+    "kilograms": "kilogram",
+    "lb": "pound",
+    "lbs": "pound",
+    "pound": "pound",
+    "pounds": "pound",
+    "m": "meter",
+    "meter": "meter",
+    "meters": "meter",
+    "cm": "centimeter",
+    "centimeter": "centimeter",
+    "centimeters": "centimeter",
+    "km": "kilometer",
+    "kilometer": "kilometer",
+    "kilometers": "kilometer",
+    "mile": "mile",
+    "miles": "mile",
+}
+
+NORMALIZATION = {
+    "gram": ("mass", 0.001),
+    "kilogram": ("mass", 1.0),
+    "pound": ("mass", 0.45359237),
+    "meter": ("length", 1.0),
+    "centimeter": ("length", 0.01),
+    "kilometer": ("length", 1000.0),
+    "mile": ("length", 1609.344),
+}
+
+
+def normalize_unit(unit):
+    canonical = UNIT_ALIASES.get((unit or "").lower())
+    return NORMALIZATION.get(canonical) if canonical else None

Halgorithem/claim_extraction.py

@@ -1,18 +1,14 @@
 import re
 
-from .nlp import nlp
+from .nlp import parse
 
 
 CLAIM_SPLIT_RE = re.compile(r"\s*(?:;|\n+|\s+-\s+)\s*")
-CONJUNCTION_RE = re.compile(
-    r"\s+(?:and|but|while|whereas)\s+"
-    r"(?=(?:[A-Z][a-z]+|\d|it\b|he\b|she\b|they\b|the\b|a\b|an\b))",
-    re.IGNORECASE,
-)
+CONJUNCTION_RE = re.compile(r"\s+(?:and|but|while|whereas)\s+", re.IGNORECASE)
 
 
 def _has_factual_shape(text):
-    doc = nlp(text)
+    doc = parse(text)
     has_subject = any(t.dep_ in {"nsubj", "nsubjpass"} for t in doc)
     has_verb = any(t.pos_ in {"VERB", "AUX"} for t in doc)
     has_anchor = any(doc.ents) or any(t.like_num for t in doc) or any(t.pos_ == "PROPN" for t in doc)
@@ -27,7 +23,7 @@ def _has_factual_shape(text):
 
 
 def _has_event_verb(text):
-    doc = nlp(text)
+    doc = parse(text)
     return any(t.pos_ in {"VERB", "AUX"} for t in doc)
 
 

Halgorithem/confidence.py

@@ -1,3 +1,6 @@
+from .nlp import parse
+
+
 INFERENTIAL_TERMS = {
     "helped",
     "made",
@@ -9,6 +12,14 @@
     "significant",
     "influential",
 }
+INFERENTIAL_ROOT_LEMMAS = {
+    "help",
+    "ease",
+    "learn",
+    "influence",
+    "matter",
+    "signify",
+}
 
 NEGATION_TERMS = {
     "no",
@@ -28,8 +39,9 @@
 
 
 def is_inferential_claim(claim):
-    words = set((claim or "").lower().replace(".", "").split())
-    return bool(words & INFERENTIAL_TERMS)
+    doc = parse(claim)
+    root = next((t for t in doc if t.dep_ == "ROOT"), None)
+    return bool(root and root.lemma_.lower() in INFERENTIAL_ROOT_LEMMAS)
 
 
 def is_negative_claim(claim):
@@ -42,7 +54,14 @@ def classify_support(score, threshold=0.30, contradiction=None, unsupported_term
     supported_threshold = max(threshold + 0.10, 0.40)
 
     hard_contradiction = contradiction and contradiction.get("reason") in {
-        "Date mismatch", "Number mismatch", "Unit mismatch", "Negation mismatch"
+        "Date mismatch",
+        "Number mismatch",
+        "Unit mismatch",
+        "Negation mismatch",
+        "Entity-role mismatch",
+        "Location mismatch",
+        "Source qualifier mismatch",
+        "NLI contradiction",
     }
     if hard_contradiction:
         return "CONTRADICTION"
@@ -71,6 +90,6 @@ def confidence_score(score, evidence_count=0, contradiction=None, unsupported_te
         confidence = max(0.0, min(float(score), 1.0))
     confidence += min(evidence_count, 3) * 0.04
     confidence -= min(len(unsupported_terms), 4) * 0.06
-    if contradiction:
+    if contradiction and status in {"CONTRADICTION", "HALLUCINATION"}:
         confidence += 0.10
     return round(max(0.0, min(confidence, 1.0)), 3)