File size: 8,529 Bytes
98e014c
 
 
 
 
 
 
 
a6df84a
98e014c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Cross-Encoder Reranking Module

Improves retrieval quality by reranking initial results
using a more powerful cross-encoder model.
"""

import logging
from typing import List, Optional
from dataclasses import dataclass
from langchain_core.documents import Document

logger = logging.getLogger(__name__)


@dataclass
class RerankResult:
    """Result of reranking operation."""
    documents: List[Document]
    scores: List[float]
    original_scores: List[float]


class CrossEncoderReranker:
    """
    Reranks documents using a cross-encoder model.
    
    Cross-encoders are more accurate than bi-encoders for similarity
    scoring because they process query and document together.
    """
    
    def __init__(
        self,
        model_name: str = "cross-encoder/ms-marco-MiniLM-L-6-v2",
        top_k: int = 5,
        batch_size: int = 32,
    ):
        """
        Initialize the reranker.
        
        Args:
            model_name: HuggingFace model name for cross-encoder
            top_k: Number of documents to return after reranking
            batch_size: Batch size for processing
        """
        self.model_name = model_name
        self.top_k = top_k
        self.batch_size = batch_size
        self._model = None
    
    @property
    def model(self):
        """Lazy load the cross-encoder model."""
        if self._model is None:
            try:
                from sentence_transformers import CrossEncoder
                self._model = CrossEncoder(self.model_name)
                logger.info(f"Loaded cross-encoder model: {self.model_name}")
            except ImportError:
                logger.error("sentence-transformers not installed for reranking")
                raise ImportError(
                    "Please install sentence-transformers: "
                    "pip install sentence-transformers"
                )
        return self._model
    
    def rerank(
        self,
        query: str,
        documents: List[Document],
        original_scores: Optional[List[float]] = None,
    ) -> RerankResult:
        """
        Rerank documents using cross-encoder.
        
        Args:
            query: The search query
            documents: List of documents to rerank
            original_scores: Original retrieval scores (optional)
        
        Returns:
            RerankResult with reranked documents and scores
        """
        if not documents:
            return RerankResult(
                documents=[],
                scores=[],
                original_scores=[],
            )
        
        if original_scores is None:
            original_scores = [0.0] * len(documents)
        
        # Create query-document pairs
        pairs = [(query, doc.page_content) for doc in documents]
        
        # Get cross-encoder scores
        try:
            scores = self.model.predict(pairs, batch_size=self.batch_size)
            scores = scores.tolist() if hasattr(scores, 'tolist') else list(scores)
        except Exception as e:
            logger.error(f"Reranking failed: {e}")
            # Fall back to original order
            return RerankResult(
                documents=documents[:self.top_k],
                scores=original_scores[:self.top_k],
                original_scores=original_scores[:self.top_k],
            )
        
        # Sort by score (descending)
        doc_score_pairs = list(zip(documents, scores, original_scores))
        doc_score_pairs.sort(key=lambda x: x[1], reverse=True)
        
        # Take top_k
        top_docs = doc_score_pairs[:self.top_k]
        
        return RerankResult(
            documents=[d[0] for d in top_docs],
            scores=[d[1] for d in top_docs],
            original_scores=[d[2] for d in top_docs],
        )


class LLMReranker:
    """
    Reranks documents using an LLM for more nuanced scoring.
    
    Uses prompt-based relevance scoring with Gemini.
    """
    
    def __init__(
        self,
        top_k: int = 5,
        model: str = "gemini-2.0-flash",
    ):
        self.top_k = top_k
        self.model = model
        self._llm = None
    
    @property
    def llm(self):
        """Lazy load the LLM."""
        if self._llm is None:
            from src.llmops.llm_manager import LLMManager
            manager = LLMManager(model=self.model)
            self._llm = manager.get_llm()
        return self._llm
    
    def rerank(
        self,
        query: str,
        documents: List[Document],
        original_scores: Optional[List[float]] = None,
    ) -> RerankResult:
        """
        Rerank using LLM-based relevance scoring.
        
        Args:
            query: The search query
            documents: List of documents to rerank
            original_scores: Original retrieval scores
        
        Returns:
            RerankResult with reranked documents
        """
        if not documents:
            return RerankResult(documents=[], scores=[], original_scores=[])
        
        if original_scores is None:
            original_scores = [0.0] * len(documents)
        
        scores = []
        
        for doc in documents:
            prompt = f"""Rate the relevance of this document to the query.
            
Query: {query}

Document: {doc.page_content[:500]}

Rate from 0-10 where 10 is highly relevant. Reply with just the number."""
            
            try:
                response = self.llm.invoke(prompt)
                score = float(response.content.strip())
                scores.append(min(max(score, 0), 10))  # Clamp to 0-10
            except Exception as e:
                logger.warning(f"LLM scoring failed: {e}")
                scores.append(5.0)  # Default mid-score
        
        # Sort by score
        doc_score_pairs = list(zip(documents, scores, original_scores))
        doc_score_pairs.sort(key=lambda x: x[1], reverse=True)
        
        top_docs = doc_score_pairs[:self.top_k]
        
        return RerankResult(
            documents=[d[0] for d in top_docs],
            scores=[d[1] for d in top_docs],
            original_scores=[d[2] for d in top_docs],
        )


class HybridReranker:
    """
    Combines cross-encoder and LLM reranking.
    
    Uses cross-encoder for speed, with optional LLM refinement
    for top candidates.
    """
    
    def __init__(
        self,
        cross_encoder_model: str = "cross-encoder/ms-marco-MiniLM-L-6-v2",
        cross_encoder_top_k: int = 10,
        final_top_k: int = 5,
        use_llm_refinement: bool = False,
    ):
        self.cross_encoder = CrossEncoderReranker(
            model_name=cross_encoder_model,
            top_k=cross_encoder_top_k,
        )
        self.final_top_k = final_top_k
        self.use_llm_refinement = use_llm_refinement
        
        if use_llm_refinement:
            self.llm_reranker = LLMReranker(top_k=final_top_k)
    
    def rerank(
        self,
        query: str,
        documents: List[Document],
        original_scores: Optional[List[float]] = None,
    ) -> RerankResult:
        """Hybrid reranking with cross-encoder and optional LLM."""
        
        # First pass: cross-encoder
        ce_result = self.cross_encoder.rerank(query, documents, original_scores)
        
        if not self.use_llm_refinement:
            # Just return cross-encoder results
            return RerankResult(
                documents=ce_result.documents[:self.final_top_k],
                scores=ce_result.scores[:self.final_top_k],
                original_scores=ce_result.original_scores[:self.final_top_k],
            )
        
        # Second pass: LLM refinement on top candidates
        llm_result = self.llm_reranker.rerank(
            query,
            ce_result.documents,
            ce_result.scores,
        )
        
        return llm_result


def create_reranker(
    reranker_type: str = "cross-encoder",
    **kwargs,
) -> CrossEncoderReranker | LLMReranker | HybridReranker:
    """
    Factory function to create a reranker.
    
    Args:
        reranker_type: One of "cross-encoder", "llm", "hybrid"
        **kwargs: Additional arguments passed to reranker
    
    Returns:
        Configured reranker instance
    """
    rerankers = {
        "cross-encoder": CrossEncoderReranker,
        "llm": LLMReranker,
        "hybrid": HybridReranker,
    }
    
    if reranker_type not in rerankers:
        raise ValueError(f"Unknown reranker type: {reranker_type}")
    
    return rerankers[reranker_type](**kwargs)