Spaces:
Running
Running
File size: 10,853 Bytes
5e1a30c |
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 |
"""
Semantic Reranker implementation for Modular Retriever Architecture.
This module provides a direct implementation of cross-encoder based reranking
to improve retrieval quality by reordering candidates based on query relevance.
"""
import logging
from typing import List, Dict, Any, Tuple, Optional
import numpy as np
from src.core.interfaces import Document
from .base import Reranker
logger = logging.getLogger(__name__)
class SemanticReranker(Reranker):
"""
Semantic reranker using cross-encoder models.
This is a direct implementation that uses cross-encoder models to rerank
retrieval results based on query-document relevance scores.
No external API dependencies are required.
Features:
- Cross-encoder model support (sentence-transformers)
- Configurable reranking threshold
- Batch processing for efficiency
- Optional reranking (can be disabled)
- Performance monitoring
Example:
config = {
"model": "cross-encoder/ms-marco-MiniLM-L-6-v2",
"enabled": True,
"batch_size": 32,
"top_k": 10,
"score_threshold": 0.0
}
reranker = SemanticReranker(config)
results = reranker.rerank(query, documents, initial_scores)
"""
def __init__(self, config: Dict[str, Any]):
"""
Initialize semantic reranker.
Args:
config: Configuration dictionary with:
- model: Cross-encoder model name (default: "cross-encoder/ms-marco-MiniLM-L-6-v2")
- enabled: Whether reranking is enabled (default: True)
- batch_size: Batch size for model inference (default: 32)
- top_k: Maximum number of documents to rerank (default: 10)
- score_threshold: Minimum score threshold for reranking (default: 0.0)
- device: Device to run model on (default: "auto")
"""
self.config = config
self.model_name = config.get("model", "cross-encoder/ms-marco-MiniLM-L-6-v2")
self.enabled = config.get("enabled", True)
self.batch_size = config.get("batch_size", 32)
self.top_k = config.get("top_k", 10)
self.score_threshold = config.get("score_threshold", 0.0)
self.device = config.get("device", "auto")
# Initialize model lazily
self.model = None
self._model_loaded = False
self._initialized = True # Always initialized (model loading is lazy)
# Validation
if self.batch_size <= 0:
raise ValueError("batch_size must be positive")
if self.top_k <= 0:
raise ValueError("top_k must be positive")
logger.info(f"SemanticReranker initialized with model={self.model_name}, enabled={self.enabled}")
def _load_model(self) -> None:
"""Load the cross-encoder model lazily."""
if self._model_loaded:
return
if not self.enabled:
logger.info("Reranker disabled, skipping model loading")
self._model_loaded = True
return
try:
from sentence_transformers import CrossEncoder
logger.info(f"Loading cross-encoder model: {self.model_name}")
self.model = CrossEncoder(self.model_name, device=self.device)
self._model_loaded = True
logger.info("Cross-encoder model loaded successfully")
except ImportError:
logger.warning("sentence-transformers not available, disabling reranker")
self.enabled = False
self._model_loaded = True
except Exception as e:
logger.error(f"Failed to load cross-encoder model: {e}")
self.enabled = False
self._model_loaded = True
def rerank(
self,
query: str,
documents: List[Document],
initial_scores: List[float]
) -> List[Tuple[int, float]]:
"""
Rerank documents based on query relevance.
Args:
query: The search query
documents: List of candidate documents
initial_scores: Initial relevance scores from fusion
Returns:
List of (document_index, reranked_score) tuples sorted by score
"""
if not self.enabled:
# Return original ranking if reranking is disabled
return [(i, score) for i, score in enumerate(initial_scores)]
if not documents or not query.strip():
return []
# Load model if not already loaded
self._load_model()
if not self._model_loaded or self.model is None:
# Fallback to original ranking
return [(i, score) for i, score in enumerate(initial_scores)]
# Limit to top_k documents for efficiency
num_docs = min(len(documents), self.top_k)
# Create query-document pairs for cross-encoder
query_doc_pairs = []
doc_indices = []
for i in range(num_docs):
doc_text = documents[i].content
# Truncate very long documents for efficiency
if len(doc_text) > 2000:
doc_text = doc_text[:2000] + "..."
query_doc_pairs.append([query, doc_text])
doc_indices.append(i)
try:
# Get cross-encoder scores in batches
cross_encoder_scores = []
for i in range(0, len(query_doc_pairs), self.batch_size):
batch = query_doc_pairs[i:i + self.batch_size]
batch_scores = self.model.predict(batch)
cross_encoder_scores.extend(batch_scores)
# Create reranked results
reranked_results = []
for i, score in enumerate(cross_encoder_scores):
doc_idx = doc_indices[i]
# Apply score threshold
if score >= self.score_threshold:
reranked_results.append((doc_idx, float(score)))
# Add remaining documents that weren't reranked
for i in range(num_docs, len(documents)):
if i < len(initial_scores):
reranked_results.append((i, initial_scores[i]))
# Sort by reranked score (descending)
reranked_results.sort(key=lambda x: x[1], reverse=True)
return reranked_results
except Exception as e:
logger.error(f"Reranking failed: {e}, falling back to original ranking")
return [(i, score) for i, score in enumerate(initial_scores)]
def is_enabled(self) -> bool:
"""
Check if reranking is enabled.
Returns:
True if reranking should be performed
"""
return self.enabled
def get_reranker_info(self) -> Dict[str, Any]:
"""
Get information about the reranker.
Returns:
Dictionary with reranker configuration and statistics
"""
info = {
"model": self.model_name,
"enabled": self.enabled,
"batch_size": self.batch_size,
"top_k": self.top_k,
"score_threshold": self.score_threshold,
"device": self.device,
"model_loaded": self._model_loaded
}
if self.model is not None:
info["model_device"] = str(self.model.device)
return info
def enable(self) -> None:
"""Enable reranking."""
self.enabled = True
logger.info("Reranker enabled")
def disable(self) -> None:
"""Disable reranking."""
self.enabled = False
logger.info("Reranker disabled")
def set_top_k(self, top_k: int) -> None:
"""
Set the maximum number of documents to rerank.
Args:
top_k: Maximum number of documents to rerank
"""
if top_k <= 0:
raise ValueError("top_k must be positive")
self.top_k = top_k
logger.info(f"Reranker top_k set to {top_k}")
def set_score_threshold(self, threshold: float) -> None:
"""
Set the minimum score threshold for reranking.
Args:
threshold: Minimum score threshold
"""
self.score_threshold = threshold
logger.info(f"Reranker score threshold set to {threshold}")
def predict_scores(self, query: str, documents: List[Document]) -> List[float]:
"""
Get cross-encoder scores for query-document pairs.
Args:
query: The search query
documents: List of documents
Returns:
List of relevance scores
"""
if not self.enabled:
return [0.0] * len(documents)
self._load_model()
if not self._model_loaded or self.model is None:
return [0.0] * len(documents)
# Create query-document pairs
query_doc_pairs = []
for doc in documents:
doc_text = doc.content
if len(doc_text) > 2000:
doc_text = doc_text[:2000] + "..."
query_doc_pairs.append([query, doc_text])
try:
# Get scores in batches
scores = []
for i in range(0, len(query_doc_pairs), self.batch_size):
batch = query_doc_pairs[i:i + self.batch_size]
batch_scores = self.model.predict(batch)
scores.extend(batch_scores)
return [float(score) for score in scores]
except Exception as e:
logger.error(f"Score prediction failed: {e}")
return [0.0] * len(documents)
def get_model_info(self) -> Dict[str, Any]:
"""
Get information about the loaded model.
Returns:
Dictionary with model information
"""
if not self._model_loaded or self.model is None:
return {"status": "not_loaded"}
info = {
"status": "loaded",
"model_name": self.model_name,
"device": str(self.model.device) if hasattr(self.model, 'device') else "unknown"
}
# Try to get model-specific info
try:
if hasattr(self.model, 'model'):
info["model_type"] = type(self.model.model).__name__
if hasattr(self.model, 'tokenizer'):
info["tokenizer_type"] = type(self.model.tokenizer).__name__
except:
pass
return info |