Spaces:
Sleeping
Sleeping
File size: 21,009 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 |
"""
BM25 Sparse Retriever implementation for Modular Retriever Architecture.
This module provides a direct implementation of BM25 sparse retrieval
extracted from the existing sparse retrieval system for improved modularity.
"""
import logging
import re
import time
from typing import List, Dict, Any, Tuple, Optional
import numpy as np
from rank_bm25 import BM25Okapi
from src.core.interfaces import Document
from .base import SparseRetriever
logger = logging.getLogger(__name__)
class BM25Retriever(SparseRetriever):
"""
BM25-based sparse retrieval implementation.
This is a direct implementation that handles BM25 keyword search
without external adapters. It provides efficient sparse retrieval
for technical documentation with optimized tokenization.
Features:
- Technical term preservation (handles RISC-V, ARM Cortex-M, etc.)
- Configurable BM25 parameters (k1, b)
- Normalized scoring for fusion compatibility
- Efficient preprocessing and indexing
- Performance monitoring
Example:
config = {
"k1": 1.2,
"b": 0.75,
"lowercase": True,
"preserve_technical_terms": True
}
retriever = BM25Retriever(config)
retriever.index_documents(documents)
results = retriever.search("RISC-V processor", k=5)
"""
def __init__(self, config: Dict[str, Any]):
"""
Initialize BM25 sparse retriever.
Args:
config: Configuration dictionary with:
- k1: Term frequency saturation parameter (default: 1.2)
- b: Document length normalization factor (default: 0.75)
- lowercase: Whether to lowercase text (default: True)
- preserve_technical_terms: Whether to preserve technical terms (default: True)
- filter_stop_words: Whether to filter common stop words (default: True)
- stop_word_sets: List of predefined stopword sets to use (default: ["english_common"])
- custom_stop_words: Additional stop words to filter (default: empty list)
- min_word_length: Minimum word length to preserve (default: 2)
- debug_stop_words: Enable debug logging for stopword filtering (default: False)
- min_score: Minimum normalized score threshold for results (default: 0.0)
"""
self.config = config
self.k1 = config.get("k1", 1.2)
self.b = config.get("b", 0.75)
self.lowercase = config.get("lowercase", True)
self.preserve_technical_terms = config.get("preserve_technical_terms", True)
self.filter_stop_words = config.get("filter_stop_words", True)
self.stop_word_sets = config.get("stop_word_sets", ["english_common"])
self.custom_stop_words = set(config.get("custom_stop_words", []))
self.min_word_length = config.get("min_word_length", 2)
self.debug_stop_words = config.get("debug_stop_words", False)
self.min_score = config.get("min_score", 0.0)
# Initialize stopword sets
self._initialize_stopword_sets()
# Combine all stopword sets
self.stop_words = set()
if self.filter_stop_words:
for set_name in self.stop_word_sets:
if set_name in self.available_stop_word_sets:
self.stop_words.update(self.available_stop_word_sets[set_name])
else:
logger.warning(f"Unknown stopword set: {set_name}")
# Add custom stop words
self.stop_words.update(self.custom_stop_words)
else:
# Only use custom stop words if filtering is disabled
self.stop_words = self.custom_stop_words.copy()
# Validation
if self.k1 <= 0:
raise ValueError("k1 must be positive")
if not 0 <= self.b <= 1:
raise ValueError("b must be between 0 and 1")
# BM25 components
self.bm25: Optional[BM25Okapi] = None
self.documents: List[Document] = []
self.tokenized_corpus: List[List[str]] = []
self.chunk_mapping: List[int] = []
# Deferred indexing control
self._index_dirty = False # Track if index needs rebuilding
self._deferred_mode = False # Enable deferred indexing mode
# Compile regex patterns for technical term preservation
if self.preserve_technical_terms:
self._tech_pattern = re.compile(r'[a-zA-Z0-9][\w\-_.]*[a-zA-Z0-9]|[a-zA-Z0-9]')
self._punctuation_pattern = re.compile(r'[^\w\s\-_.]')
else:
self._tech_pattern = re.compile(r'\b\w+\b')
self._punctuation_pattern = re.compile(r'[^\w\s]')
logger.info(f"BM25Retriever initialized with k1={self.k1}, b={self.b}, stop_word_sets={self.stop_word_sets}, stop_words={len(self.stop_words)}")
def _initialize_stopword_sets(self) -> None:
"""
Initialize predefined stopword sets for different filtering strategies.
"""
# Standard English stop words (articles, prepositions, common verbs)
english_common = {
'a', 'an', 'and', 'are', 'as', 'at', 'be', 'by', 'for', 'from', 'has', 'he', 'in', 'is', 'it',
'its', 'of', 'on', 'that', 'the', 'to', 'was', 'were', 'will', 'with', 'this', 'but',
'they', 'have', 'had', 'what', 'said', 'each', 'which', 'she', 'do', 'how', 'their', 'if',
'up', 'out', 'many', 'then', 'them', 'these', 'so', 'some', 'her', 'would', 'make', 'like',
'into', 'him', 'time', 'two', 'more', 'go', 'no', 'way', 'could', 'my', 'than', 'first',
'been', 'call', 'who', 'sit', 'now', 'find', 'down', 'day', 'did', 'get', 'come',
'made', 'may', 'part', 'much', 'too', 'any', 'after', 'back', 'other', 'see',
'want', 'just', 'also', 'when', 'here', 'all', 'well', 'can', 'should', 'must', 'might',
'shall', 'about', 'before', 'through', 'over', 'under', 'above', 'below', 'between', 'among'
}
# NOTE: Removed interrogative_words and irrelevant_entities sets
# These contained discriminative terms that should be preserved for proper BM25 behavior
# BM25 is designed for lexical matching only, not semantic analysis
# Extended set for comprehensive filtering
english_extended = english_common | {
'very', 'quite', 'really', 'actually', 'basically', 'essentially', 'generally',
'specifically', 'particularly', 'especially', 'exactly', 'precisely', 'approximately',
'roughly', 'mostly', 'mainly', 'primarily', 'largely', 'completely', 'totally',
'absolutely', 'definitely', 'certainly', 'probably', 'possibly', 'perhaps',
'maybe', 'sometimes', 'often', 'usually', 'always', 'never', 'rarely', 'seldom',
'frequently', 'occasionally', 'constantly', 'continuously', 'immediately', 'suddenly',
'quickly', 'slowly', 'carefully', 'easily', 'simply', 'clearly', 'obviously'
}
# Minimal set for technical domains (preserves more terms)
technical_minimal = {
'a', 'an', 'the', 'and', 'or', 'but', 'in', 'on', 'at', 'to', 'for', 'of', 'with',
'by', 'from', 'up', 'out', 'down', 'off', 'over', 'under', 'again', 'further',
'then', 'once', 'here', 'there', 'when', 'where', 'why', 'how', 'all', 'any',
'both', 'each', 'few', 'more', 'most', 'other', 'some', 'such', 'no', 'nor',
'not', 'only', 'own', 'same', 'so', 'than', 'too', 'very'
}
self.available_stop_word_sets = {
"english_common": english_common,
"english_extended": english_extended,
"technical_minimal": technical_minimal
}
def enable_deferred_indexing(self) -> None:
"""Enable deferred indexing mode to avoid rebuilding index on every batch"""
self._deferred_mode = True
logger.debug("Deferred indexing mode enabled")
def disable_deferred_indexing(self) -> None:
"""Disable deferred indexing mode and rebuild index if needed"""
self._deferred_mode = False
if self._index_dirty:
self._rebuild_index()
logger.debug("Deferred indexing mode disabled")
def force_rebuild_index(self) -> None:
"""Force rebuild the BM25 index with all accumulated documents"""
if self.tokenized_corpus:
self._rebuild_index()
else:
logger.warning("No documents to rebuild index")
def _rebuild_index(self) -> None:
"""Internal method to rebuild the BM25 index"""
if not self.tokenized_corpus:
logger.warning("No tokenized corpus available for index rebuild")
return
start_time = time.time()
self.bm25 = BM25Okapi(self.tokenized_corpus, k1=self.k1, b=self.b)
self._index_dirty = False
elapsed = time.time() - start_time
total_tokens = sum(len(tokens) for tokens in self.tokenized_corpus)
valid_doc_count = len([tokens for tokens in self.tokenized_corpus if tokens])
logger.info(f"Rebuilt BM25 index with {valid_doc_count} documents in {elapsed:.3f}s")
def index_documents(self, documents: List[Document]) -> None:
"""
Index documents for BM25 sparse retrieval.
Args:
documents: List of documents to index
"""
if not documents:
raise ValueError("Cannot index empty document list")
start_time = time.time()
# Store documents (extend existing instead of replacing)
if not hasattr(self, 'documents') or self.documents is None:
self.documents = []
if not hasattr(self, 'tokenized_corpus') or self.tokenized_corpus is None:
self.tokenized_corpus = []
if not hasattr(self, 'chunk_mapping') or self.chunk_mapping is None:
self.chunk_mapping = []
# Keep track of starting index for new documents
start_idx = len(self.documents)
# Add new documents
self.documents.extend(documents)
# Extract and preprocess texts for new documents only
texts = [doc.content for doc in documents]
new_tokenized = [self._preprocess_text(text) for text in texts]
# Filter out empty tokenized texts and track mapping for new documents
for i, tokens in enumerate(new_tokenized):
if tokens: # Only include non-empty tokenized texts
self.tokenized_corpus.append(tokens)
self.chunk_mapping.append(start_idx + i)
if not self.tokenized_corpus:
raise ValueError("No valid text content found in documents")
# Rebuild BM25 index unless in deferred mode
if self._deferred_mode:
# Mark index as dirty but don't rebuild yet
self._index_dirty = True
logger.debug(f"Added {len(documents)} documents to corpus (deferred mode - index not rebuilt)")
else:
# Rebuild index immediately (original behavior)
self._rebuild_index()
elapsed = time.time() - start_time
total_tokens = sum(len(tokens) for tokens in self.tokenized_corpus)
tokens_per_sec = total_tokens / elapsed if elapsed > 0 else 0
valid_doc_count = len([tokens for tokens in self.tokenized_corpus if tokens])
logger.info(f"Indexed {len(documents)} new documents ({valid_doc_count} total valid) in {elapsed:.3f}s")
logger.debug(f"Processing rate: {tokens_per_sec:.1f} tokens/second")
def search(self, query: str, k: int = 5) -> List[Tuple[int, float]]:
"""
Search for documents using BM25 sparse retrieval.
Args:
query: Search query string
k: Number of results to return
Returns:
List of (document_index, score) tuples sorted by relevance
"""
# Ensure index is built before searching
if self.bm25 is None or self._index_dirty:
if self._index_dirty:
logger.debug("Rebuilding BM25 index before search (was dirty)")
self._rebuild_index()
else:
raise ValueError("Must call index_documents() before searching")
if not query or not query.strip():
return []
if k <= 0:
raise ValueError("k must be positive")
# Preprocess query using same method as documents
query_tokens = self._preprocess_text(query)
if not query_tokens:
return []
# Get BM25 scores for all documents
scores = self.bm25.get_scores(query_tokens)
if len(scores) == 0:
return []
# Fix for rank_bm25 library bug: BM25 scores can be negative when they shouldn't be
# Ensure all scores are non-negative by shifting them if needed
min_raw_score = np.min(scores)
if min_raw_score < 0:
scores = scores - min_raw_score # Shift all scores to be non-negative
logger.debug(f"Shifted negative BM25 scores by {-min_raw_score:.6f}")
# Normalize scores to [0,1] range for fusion compatibility
max_score = np.max(scores)
min_score = np.min(scores)
if max_score > min_score:
# Standard min-max normalization to [0,1]
normalized_scores = (scores - min_score) / (max_score - min_score)
else:
# All scores are the same - check if any actual matches exist
if np.any(scores != 0):
# Scores are equal and non-zero (all docs equally relevant)
normalized_scores = np.ones_like(scores)
else:
# All scores are exactly zero (no matches)
normalized_scores = np.zeros_like(scores)
# Create results with original document indices
results = [
(self.chunk_mapping[i], float(normalized_scores[i]))
for i in range(len(scores))
]
# Filter out zero scores (no matches) and apply minimum score threshold
threshold = max(self.min_score, 0.001) # Always filter scores <= 0
filtered_results = [(doc_idx, score) for doc_idx, score in results if score >= threshold]
if not filtered_results:
logger.debug(f"No BM25 results above score threshold {threshold}")
return []
results = filtered_results
# Sort by score (descending) and return top_k
results.sort(key=lambda x: x[1], reverse=True)
return results[:k]
def get_document_count(self) -> int:
"""Get the number of indexed documents."""
return len(self.documents)
def clear(self) -> None:
"""Clear all indexed documents."""
self.documents.clear()
self.tokenized_corpus.clear()
self.chunk_mapping.clear()
self.bm25 = None
logger.info("BM25 index cleared")
def get_stats(self) -> Dict[str, Any]:
"""
Get statistics about the BM25 retriever.
Returns:
Dictionary with retriever statistics
"""
stats = {
"k1": self.k1,
"b": self.b,
"lowercase": self.lowercase,
"preserve_technical_terms": self.preserve_technical_terms,
"filter_stop_words": self.filter_stop_words,
"stop_word_sets": self.stop_word_sets,
"stop_words_count": len(self.stop_words) if self.stop_words else 0,
"min_word_length": self.min_word_length,
"debug_stop_words": self.debug_stop_words,
"min_score": self.min_score,
"total_documents": len(self.documents),
"valid_documents": len(self.chunk_mapping),
"is_indexed": self.bm25 is not None
}
if self.tokenized_corpus:
total_tokens = sum(len(tokens) for tokens in self.tokenized_corpus)
stats.update({
"total_tokens": total_tokens,
"avg_tokens_per_doc": total_tokens / len(self.tokenized_corpus) if self.tokenized_corpus else 0
})
return stats
def _preprocess_text(self, text: str) -> List[str]:
"""
Preprocess text with standard BM25 stopword filtering.
Args:
text: Raw text to tokenize
Returns:
List of preprocessed tokens
"""
if not text or not text.strip():
return []
original_text = text
# Convert to lowercase while preserving structure
if self.lowercase:
text = text.lower()
# Remove punctuation except hyphens, underscores, periods in technical terms
text = self._punctuation_pattern.sub(' ', text)
# Extract tokens using appropriate pattern
tokens = self._tech_pattern.findall(text)
# Filter out tokens shorter than minimum length
if self.min_word_length > 1:
length_filtered = [token for token in tokens if len(token) >= self.min_word_length]
else:
length_filtered = [token for token in tokens if len(token) > 0]
# Apply standard stopword filtering (linguistic noise words only)
if self.stop_words:
filtered_tokens = []
stop_words_removed = []
for token in length_filtered:
token_lower = token.lower()
# Simple standard stopword filtering - no semantic analysis
if token_lower in self.stop_words:
stop_words_removed.append(token)
else:
filtered_tokens.append(token)
# Debug logging if enabled
if self.debug_stop_words and stop_words_removed:
logger.info(f"[BM25_DEBUG] Text: \"{original_text[:50]}{'...' if len(original_text) > 50 else ''}\"")
logger.info(f"[BM25_DEBUG] Tokens before filtering: {length_filtered}")
logger.info(f"[BM25_DEBUG] Stop words removed: {stop_words_removed}")
logger.info(f"[BM25_DEBUG] Tokens after filtering: {filtered_tokens}")
if length_filtered:
filter_rate = len(stop_words_removed) / len(length_filtered) * 100
logger.info(f"[BM25_DEBUG] Filtering impact: {filter_rate:.1f}% tokens removed")
logger.info(f"[BM25_DEBUG] ---")
return filtered_tokens
else:
return length_filtered
def get_query_tokens(self, query: str) -> List[str]:
"""
Get preprocessed tokens for a query (useful for debugging).
Args:
query: Query string
Returns:
List of preprocessed tokens
"""
return self._preprocess_text(query)
def get_document_tokens(self, doc_index: int) -> List[str]:
"""
Get preprocessed tokens for a document (useful for debugging).
Args:
doc_index: Document index
Returns:
List of preprocessed tokens
"""
if 0 <= doc_index < len(self.tokenized_corpus):
return self.tokenized_corpus[doc_index]
else:
raise IndexError(f"Document index {doc_index} out of range")
def get_bm25_scores(self, query: str) -> List[float]:
"""
Get raw BM25 scores for all documents (useful for debugging).
Args:
query: Query string
Returns:
List of BM25 scores (not normalized)
"""
if self.bm25 is None:
raise ValueError("Must call index_documents() before getting scores")
query_tokens = self._preprocess_text(query)
if not query_tokens:
return []
scores = self.bm25.get_scores(query_tokens)
return scores.tolist()
def get_term_frequencies(self, query: str) -> Dict[str, int]:
"""
Get term frequencies for a query (useful for analysis).
Args:
query: Query string
Returns:
Dictionary mapping terms to frequencies
"""
query_tokens = self._preprocess_text(query)
term_freqs = {}
for token in query_tokens:
term_freqs[token] = term_freqs.get(token, 0) + 1
return term_freqs |