File size: 11,690 Bytes
11cc0d3
 
 
 
 
 
 
 
b2ed798
 
 
11cc0d3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
91ebeb2
11cc0d3
 
 
 
 
 
 
 
b2ed798
 
 
149817f
 
b2ed798
 
 
 
 
 
149817f
 
 
 
64e7d5a
b2ed798
11cc0d3
b2ed798
11cc0d3
 
 
 
 
 
 
b2ed798
11cc0d3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b2ed798
11cc0d3
 
 
 
149817f
 
11cc0d3
 
 
 
 
 
 
 
 
 
 
 
 
 
149817f
11cc0d3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b2ed798
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import chromadb
from chromadb.utils import embedding_functions
import json
import re
from openai import OpenAI
import re
import json
import spacy
from transformers import T5ForConditionalGeneration, T5Tokenizer
from transformers import pipeline  

def robust_json_extractor(response_content):
    # Preprocess: Remove markdown code blocks and extra whitespace
    cleaned = re.sub(r'``````', '', response_content).strip()
    
    # Key-specific regex patterns
    patterns = {
        "verdict": r'"verdict"\s*:\s*"((?:\\"|[^"])*)"',
        "evidence": r'"evidence"\s*:\s*(\[[^\]]*?\]|\[.*?\])(?=\s*[,}])',
        "reasoning": r'"reasoning"\s*:\s*"((?:\\"|[^"])*)"'
    }
    
    result = {}
    for key, pattern in patterns.items():
        match = re.search(pattern, cleaned, re.DOTALL)
        if match:
            try:
                if key == "evidence":
                    # Handle array parsing with json.loads
                    evidence_str = re.sub(r'(?<!\\)"', r'\"', match.group(1))  # Escape unescaped quotes
                    result[key] = json.loads(evidence_str)
                else:
                    # Unescape quotes for strings
                    result[key] = json.loads(f'"{match.group(1)}"')  
            except:
                # Fallback: Return raw matched string
                result[key] = match.group(1)
    
    # Validation
    required_keys = ["verdict", "evidence", "reasoning"]
    if all(key in result for key in required_keys):
        return result
    else:
        # Fallback to standard JSON parsing
        try:
            return json.loads(re.search(r'\{.*\}', cleaned, re.DOTALL).group())
        except:
            return {"error": "Failed to extract required keys", "raw": cleaned}

class FactChecker:
    def __init__(self, chroma_path, collection_name, groq_client):
        self.client = chromadb.PersistentClient(path=chroma_path)
        self.collection = self.client.get_collection(
            name=collection_name,
            embedding_function=embedding_functions.SentenceTransformerEmbeddingFunction(
                model_name="all-MiniLM-L6-v2"
            )
        )
        self.groq_client = groq_client
        self.model_name = "llama3-8b-8192"
        self.ner = spacy.load("en_core_web_sm")
        

        # self.claim_tokenizer = T5Tokenizer.from_pretrained("Babelscape/t5-base-summarization-claim-extractor")
        # self.claim_model = T5ForConditionalGeneration.from_pretrained("Babelscape/t5-base-summarization-claim-extractor")

    def extract_entities(self, text):
        doc = self.ner(text)
        return [(ent.text, ent.label_) for ent in doc.ents]

    def extract_claims(self, text, threshold=0.5):
        # tok_input = self.claim_tokenizer.batch_encode_plus([text], return_tensors="pt", padding=True)
        # outputs = self.claim_model.generate(**tok_input)
        # claims = self.claim_tokenizer.batch_decode(outputs, skip_special_tokens=True)
        # claims = [claim.strip() for claim in claims if len(claim.strip()) > 0]
        return [text]


    def verify_single_claim(self, claim, confidence_threshold=0.5):
        results = self.collection.query(
            query_texts=[claim],
            n_results=3,
            include=["documents", "metadatas", "distances"]
        )
        zipped_results = sorted(
            zip(results['documents'][0], results['metadatas'][0], results['distances'][0]),
            key=lambda x: x[2]
        )
        evidence = []
        for doc, meta, distance in zipped_results:
            source = meta["source"] if meta and "source" in meta else "Unknown source"
            similarity_score = 1 - (distance / 2)  # Assuming cosine distance in [0,2]
            evidence.append(
                f'"{doc}" (Source: {source}, Similarity: {similarity_score:.2f})'
            )
        avg_distance = sum(d for _, _, d in zipped_results) / len(zipped_results)
        confidence = 1 - (avg_distance / 2)  # Normalize to 0-1 range

        if confidence < confidence_threshold:
            return {
                "verdict": "Unverifiable",
                "confidence": confidence,
                "evidence": [e.split(" (Source:")[0] for e in evidence],
                "reasoning": "Claim is too vague or lacks sufficient evidence"
            }

        evidence_str = "\n".join([f"- {e}" for e in evidence])
        prompt = f""" You are a powerful fact checker. Analyze the claim below against the provided verified information. 
Relying on the similarity scores, also carefully check whether all factual details in the claim (such as dates, names, locations, and events) exactly match atleast one of the evidence. If from first evidence, evidence is not sufficient, use the next evidence to verify the claim. 
If there is any factual mismatch (for example, the date in the claim is different from the evidence), classify the claim as False. Any factual mismatch, even if the overall context is similar, should lead to a False classification.
If the evidence is too vague or lacks strong matches, classify as Unverifiable.
If evidence directly contradicts the claim, classify as False.
Any discrepancy in factual details, even if the overall context is similar, should lead to a False classification.
If the evidence fully supports the claim with all factual details matching, classify as True.

Claim:
{claim}

Evidence (with similarity scores):
{evidence_str}

Guidelines:
1. Give more weight to evidence with higher similarity scores, but do not ignore factual mismatches.
2. If any one piece of evidence independently supports the claim, without factual mismatches, classify as True.
2. Pay close attention to details such as dates, names, locations, and events.
3. If the claim and evidence differ on any factual point, do not classify as True.
4. Respond only in JSON format without any additional text.
5. In the "evidence" array, include only full evidence statements as strings, without any extra comments or explanations.
6. Put all explanations or comparisons in the "reasoning" field.

Respond in JSON format:
{{
    "verdict": "Verdict",
    "evidence": [List of relevant facts from provided evidence],
    "reasoning": "Explanation of the verdict based on evidence and factual details"
}}
"""
        completion = self.groq_client.chat.completions.create(
            model=self.model_name,
            messages=[{"role": "user", "content": prompt}],
            temperature=0.1,
            max_tokens=400
        )
        response_content = completion.choices[0].message.content
        parsed = robust_json_extractor(response_content)
        if "error" in parsed:
            return {
                "error": parsed["error"],
                "confidence": confidence,
                "raw_response": parsed.get("raw", response_content)
            }
        else:
            required_keys = ["verdict", "evidence", "reasoning"]
            if all(key in parsed for key in required_keys):
                return {
                    "verdict": parsed["verdict"],
                    "confidence": confidence,
                    "evidence": [e.split(" (Source:")[0] for e in evidence],
                    "reasoning": parsed["reasoning"]
                }
            else:
                return {
                    "error": f"Missing required keys: {[k for k in required_keys if k not in parsed]}",
                    "confidence": confidence,
                    "raw_response": response_content
                }

    def verify_single_entity(self, entity_text, confidence_threshold=0.5):
        """Verify a single named entity against the fact database"""
        # Vector similarity search
        results = self.collection.query(
            query_texts=[entity_text],
            n_results=3,
            include=["documents", "metadatas", "distances"]
        )
        
        # Process evidence with similarity normalization
        evidence = []
        total_distance = 0
        for doc, meta, distance in zip(results['documents'][0], 
                                    results['metadatas'][0], 
                                    results['distances'][0]):
            similarity = 1 - (distance / 2)  # Convert cosine distance to similarity
            evidence.append({
                "text": doc,
                "source": meta.get("source", "Unknown"),
                "similarity": similarity
            })
            total_distance += distance
        
        avg_similarity = 1 - (total_distance / len(results['distances'][0]) / 2)
        
        # Prepare LLM verification prompt
        evidence_str = "\n".join([
            f"- {e['text']} (Similarity: {e['similarity']:.2f})" 
            for e in evidence
        ])
        
        prompt = f"""**Entity Verification Task**
    Entity: "{entity_text}"

    **Verified Evidence:**
    {evidence_str}

    **Instructions:**
    1. Verify if this entity exists in official records
    2. Check for exact matches of names/titles
    3. Confirm associated details (locations, dates, roles)
    4. Return JSON with: verdict (True/False/Unverified), confidence (0-1), reasoning

    **JSON Response:"""
        
        try:
            response = self.groq_client.chat.completions.create(
                model=self.model_name,
                messages=[{"role": "user", "content": prompt}],
                temperature=0.2,
                response_format={"type": "json_object"}
            )
            
            result = json.loads(response.choices[0].message.content)
            return {
                "verdict": result.get("verdict", "Unverified"),
                "confidence": min(max(result.get("confidence", avg_similarity), 0), 1),
                "evidence": [e["text"] for e in evidence],
                "reasoning": result.get("reasoning", "No reasoning provided")
            }
            
        except Exception as e:
            return {
                "verdict": "Error",
                "confidence": 0,
                "evidence": [],
                "reasoning": f"Verification failed: {str(e)}"
            }

    def verify_claim(self, text, confidence_threshold=0.5):
        """
        Main method: takes input text, extracts entities and claims, 
        verifies each, and returns JSON results
        """
        # Extract entities and claims
        entities = self.extract_entities(text)
        claims = self.extract_claims(text)
        
        # Verify claims
        claim_results = []
        for claim in claims:
            verification = self.verify_single_claim(claim, confidence_threshold)
            claim_results.append({
                "claim": claim,
                "verdict": verification.get("verdict", "Error"),
                "confidence": verification.get("confidence", 0),
                "evidence": verification.get("evidence", []),
                "reasoning": verification.get("reasoning", "Analysis failed")
            })
        
        # Verify entities
        entity_results = []
        for entity_text, entity_label in entities:
            verification = self.verify_single_entity(entity_text, confidence_threshold)
            entity_results.append({
                "entity": entity_text,
                "type": entity_label,
                "verdict": verification.get("verdict", "Error"),
                "confidence": verification.get("confidence", 0),
                "evidence": verification.get("evidence", []),
                "reasoning": verification.get("reasoning", "Analysis failed")
            })
        
        return {
            "entities": entity_results,
            "claims": claim_results
        }