Spaces:
Running
Running
File size: 39,440 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 |
"""
Modular Unified Retriever for Architecture Compliance.
This module provides a modular implementation of the unified retriever
that decomposes functionality into well-defined sub-components following
the architecture specification.
"""
import logging
import time
from typing import List, Dict, Any, Optional, Tuple
import numpy as np
from src.core.interfaces import Retriever, Document, RetrievalResult, Embedder, HealthStatus
# Forward declaration to avoid circular import
from typing import TYPE_CHECKING
if TYPE_CHECKING:
from src.core.platform_orchestrator import PlatformOrchestrator
from .indices.base import VectorIndex
from .indices.faiss_index import FAISSIndex
from .indices.weaviate_index import WeaviateIndex
from .sparse.base import SparseRetriever
from .sparse.bm25_retriever import BM25Retriever
from .fusion.base import FusionStrategy
from .fusion.rrf_fusion import RRFFusion
from .fusion.weighted_fusion import WeightedFusion
from .fusion.graph_enhanced_fusion import GraphEnhancedRRFFusion
from .fusion.score_aware_fusion import ScoreAwareFusion
from .rerankers.base import Reranker
from .rerankers.semantic_reranker import SemanticReranker
from .rerankers.identity_reranker import IdentityReranker
from .rerankers.neural_reranker import NeuralReranker
logger = logging.getLogger(__name__)
class ModularUnifiedRetriever(Retriever):
"""
Modular unified retriever with pluggable sub-components.
This implementation follows the architecture specification by decomposing
the retrieval functionality into four distinct sub-components:
- Vector Index: Handles dense semantic search
- Sparse Retriever: Handles keyword-based search
- Fusion Strategy: Combines dense and sparse results
- Reranker: Improves result relevance
Each sub-component can be independently configured and tested,
improving modularity and maintainability.
Features:
- Architecture-compliant modular design
- Configurable sub-components
- Component factory integration
- Performance monitoring
- Backward compatibility with UnifiedRetriever
- Enhanced logging and debugging
Example:
config = {
"vector_index": {
"type": "faiss",
"config": {"index_type": "IndexFlatIP", "normalize_embeddings": True}
},
"sparse": {
"type": "bm25",
"config": {"k1": 1.2, "b": 0.75}
},
"fusion": {
"type": "rrf",
"config": {"k": 60, "weights": {"dense": 0.7, "sparse": 0.3}}
},
"reranker": {
"type": "semantic",
"config": {"enabled": True, "model": "cross-encoder/ms-marco-MiniLM-L-6-v2"}
}
}
retriever = ModularUnifiedRetriever(config, embedder)
"""
def __init__(self, config: Dict[str, Any], embedder: Embedder):
"""
Initialize the modular unified retriever.
Args:
config: Configuration dictionary with sub-component specifications
embedder: Embedder component for query encoding
"""
self.config = config
self.embedder = embedder
self.documents: List[Document] = []
# Composite filtering configuration (NEW)
composite_config = config.get("composite_filtering", {})
self.composite_filtering_enabled = composite_config.get("enabled", False)
self.fusion_weight = composite_config.get("fusion_weight", 0.7)
self.semantic_weight = composite_config.get("semantic_weight", 0.3)
self.min_composite_score = composite_config.get("min_composite_score", 0.4)
self.max_candidates_multiplier = composite_config.get("max_candidates", 15) / 10.0 # Convert to multiplier (1.5x)
# Legacy semantic gap detection configuration (DEPRECATED)
self.min_semantic_alignment = config.get("min_semantic_alignment", 0.3)
# Initialize sub-components
self.vector_index = self._create_vector_index(config.get("vector_index", {}))
self.sparse_retriever = self._create_sparse_retriever(config.get("sparse", {}))
self.fusion_strategy = self._create_fusion_strategy(config.get("fusion", {}))
self.reranker = self._create_reranker(config.get("reranker", {}))
# Performance tracking
self.retrieval_stats = {
"total_retrievals": 0,
"total_time": 0.0,
"avg_time": 0.0,
"last_retrieval_time": 0.0
}
# Backend management (for multi-backend performance testing)
self.active_backend_name = "faiss" # Default backend
self.available_backends = ["faiss", "weaviate"]
self.backend_switch_count = 0
# Platform services (initialized via initialize_services)
self.platform: Optional['PlatformOrchestrator'] = None
logger.info("ModularUnifiedRetriever initialized with all sub-components")
def _create_vector_index(self, config: Dict[str, Any]) -> VectorIndex:
"""Create vector index sub-component."""
index_type = config.get("type", "faiss")
index_config = config.get("config", {})
if index_type == "faiss":
return FAISSIndex(index_config)
elif index_type == "weaviate":
return WeaviateIndex(index_config)
else:
raise ValueError(f"Unknown vector index type: {index_type}")
def _create_sparse_retriever(self, config: Dict[str, Any]) -> SparseRetriever:
"""Create sparse retriever sub-component."""
retriever_type = config.get("type", "bm25")
retriever_config = config.get("config", {})
if retriever_type == "bm25":
return BM25Retriever(retriever_config)
else:
raise ValueError(f"Unknown sparse retriever type: {retriever_type}")
def _create_fusion_strategy(self, config: Dict[str, Any]) -> FusionStrategy:
"""Create fusion strategy sub-component."""
fusion_type = config.get("type", "rrf")
fusion_config = config.get("config", {})
if fusion_type == "rrf":
return RRFFusion(fusion_config)
elif fusion_type == "weighted":
return WeightedFusion(fusion_config)
elif fusion_type == "graph_enhanced_rrf":
return GraphEnhancedRRFFusion(fusion_config)
elif fusion_type == "score_aware":
return ScoreAwareFusion(fusion_config)
else:
raise ValueError(f"Unknown fusion strategy type: {fusion_type}. Available options: rrf, weighted, graph_enhanced_rrf, score_aware")
def _create_reranker(self, config: Dict[str, Any]) -> Reranker:
"""Create reranker sub-component."""
reranker_type = config.get("type", "identity")
reranker_config = config.get("config", {})
logger.info(f"π§ Creating reranker: type={reranker_type}, config keys={list(reranker_config.keys())}")
if reranker_type == "semantic":
reranker = SemanticReranker(reranker_config)
elif reranker_type == "identity":
reranker = IdentityReranker(reranker_config)
elif reranker_type == "neural":
try:
reranker = NeuralReranker(reranker_config)
logger.info(f"β
NeuralReranker created successfully: enabled={reranker.enabled}, initialized={reranker._initialized}")
except Exception as e:
logger.error(f"β Failed to create NeuralReranker: {e}")
logger.warning("Falling back to IdentityReranker")
reranker = IdentityReranker({"enabled": True})
else:
raise ValueError(f"Unknown reranker type: {reranker_type}")
return reranker
def retrieve(self, query: str, k: int = 5) -> List[RetrievalResult]:
"""
Retrieve relevant documents using modular hybrid search.
This method orchestrates the complete retrieval pipeline:
1. Generate query embeddings
2. Perform dense vector search
3. Perform sparse keyword search
4. Fuse results using configured strategy
5. Apply reranking if enabled
Args:
query: Search query string
k: Number of results to return
Returns:
List of retrieval results sorted by relevance score
"""
start_time = time.time()
try:
# Validation
if k <= 0:
raise ValueError("k must be positive")
if not query.strip():
raise ValueError("Query cannot be empty")
if not self.documents:
raise RuntimeError("No documents have been indexed")
logger.info(f"π MODULAR RETRIEVER: Starting retrieval for query: '{query}' (k={k})")
logger.info(f"π CORPUS STATUS: {len(self.documents)} documents indexed")
# Step 1: Generate query embeddings
query_embedding = np.array(self.embedder.embed([query])[0])
logger.info(f"π€ QUERY EMBEDDING: Generated {query_embedding.shape} dimensional vector")
# Step 2: Dense vector search (with efficiency optimization)
candidate_multiplier = int(self.max_candidates_multiplier * k) if self.composite_filtering_enabled else k*2
logger.info(f"π― DENSE SEARCH: Searching for top {candidate_multiplier} candidates")
dense_results = self.vector_index.search(query_embedding, k=candidate_multiplier)
logger.info(f"β
DENSE RESULTS: {len(dense_results)} documents found")
# Log top dense results with scores
if dense_results:
logger.info(f"π TOP DENSE SCORES:")
for i, (doc_idx, score) in enumerate(dense_results[:3]):
if doc_idx < len(self.documents):
doc_title = self.documents[doc_idx].metadata.get('title', f'doc_{doc_idx}')[:50]
logger.info(f" {i+1}. [{doc_idx}] {doc_title}... β {score:.4f}")
else:
logger.warning(f"β οΈ DENSE SEARCH: No results found!")
# Step 3: Sparse keyword search (with efficiency optimization)
logger.info(f"π SPARSE SEARCH: BM25 keyword search for '{query}' (k={candidate_multiplier})")
sparse_results = self.sparse_retriever.search(query, k=candidate_multiplier)
logger.info(f"β
SPARSE RESULTS: {len(sparse_results)} documents found")
# Log top sparse results with scores
if sparse_results:
logger.info(f"π TOP SPARSE SCORES:")
for i, (doc_idx, score) in enumerate(sparse_results[:3]):
if doc_idx < len(self.documents):
doc_title = self.documents[doc_idx].metadata.get('title', f'doc_{doc_idx}')[:50]
logger.info(f" {i+1}. [{doc_idx}] {doc_title}... β {score:.4f}")
else:
logger.warning(f"β οΈ SPARSE SEARCH: No results found!")
# Step 3.5: Set documents and query for graph enhancement (if supported)
if hasattr(self.fusion_strategy, 'set_documents_and_query'):
self.fusion_strategy.set_documents_and_query(self.documents, query)
# Step 4: Fuse results
fusion_name = self.fusion_strategy.__class__.__name__
logger.info(f"π FUSION STRATEGY: Using {fusion_name} to combine results")
fused_results = self.fusion_strategy.fuse_results(dense_results, sparse_results)
logger.info(f"β
FUSION RESULTS: {len(fused_results)} documents after fusion")
# Log top fused results with scores
if fused_results:
logger.info(f"π TOP FUSED SCORES:")
for i, (doc_idx, score) in enumerate(fused_results[:5]):
if doc_idx < len(self.documents):
doc_title = self.documents[doc_idx].metadata.get('title', f'doc_{doc_idx}')[:50]
logger.info(f" {i+1}. [{doc_idx}] {doc_title}... β {score:.4f}")
else:
logger.warning(f"β οΈ FUSION: No results after fusion!")
# Step 4.5: Composite filtering (NEW) or semantic gap detection (LEGACY)
if self.composite_filtering_enabled:
# NEW: Individual document composite scoring
filtered_results = self._calculate_composite_scores(query_embedding, fused_results)
if not filtered_results:
logger.info("Composite filtering: No documents passed quality threshold")
return []
fused_results = filtered_results # Use filtered results for reranking
else:
# LEGACY: Global semantic gap detection (DEPRECATED)
if fused_results and self.min_semantic_alignment > 0:
semantic_alignment = self._calculate_semantic_alignment(query_embedding, fused_results[:5])
if semantic_alignment < self.min_semantic_alignment:
logger.info(f"Query-document semantic alignment too low: {semantic_alignment:.3f} < {self.min_semantic_alignment}")
return [] # No semantically relevant documents found
# Step 5: Apply reranking if enabled
if self.reranker.is_enabled() and fused_results:
reranker_name = self.reranker.__class__.__name__
logger.info(f"π§ RERANKING: Using {reranker_name} to improve relevance")
# Prepare documents and scores for reranking
top_candidates = fused_results[:k*2] # Rerank top candidates
candidate_documents = [self.documents[idx] for idx, _ in top_candidates]
candidate_scores = [score for _, score in top_candidates]
logger.info(f"π RERANKING: Processing {len(top_candidates)} candidates")
reranked_results = self.reranker.rerank(query, candidate_documents, candidate_scores)
# Update final results with reranked scores
final_results = []
for local_idx, reranked_score in reranked_results:
if local_idx < len(top_candidates):
original_idx = top_candidates[local_idx][0]
final_results.append((original_idx, reranked_score))
# Add remaining documents that weren't reranked
reranked_indices = {top_candidates[local_idx][0] for local_idx, _ in reranked_results
if local_idx < len(top_candidates)}
for doc_idx, score in fused_results:
if doc_idx not in reranked_indices:
final_results.append((doc_idx, score))
# Sort by final score and limit to k
final_results.sort(key=lambda x: x[1], reverse=True)
final_results = final_results[:k]
logger.info(f"β
RERANKING: Final {len(final_results)} results after reranking")
else:
# No reranking, use fused results directly
logger.info(f"βοΈ RERANKING: Skipped (reranker disabled or no results)")
final_results = fused_results[:k]
# Convert to RetrievalResult objects
retrieval_results = []
for doc_idx, score in final_results:
if doc_idx < len(self.documents):
document = self.documents[doc_idx]
retrieval_result = RetrievalResult(
document=document,
score=float(score),
retrieval_method="modular_unified_hybrid"
)
retrieval_results.append(retrieval_result)
# Log final results summary
logger.info(f"π― FINAL RETRIEVAL RESULTS: {len(retrieval_results)} documents")
if retrieval_results:
logger.info(f"π FINAL RANKING:")
for i, result in enumerate(retrieval_results):
doc_title = result.document.metadata.get('title', f'doc_{result.document.content[:30]}')[:50]
logger.info(f" {i+1}. {doc_title}... β {result.score:.4f}")
else:
logger.warning(f"β NO RESULTS: Query '{query}' returned no relevant documents!")
# Update performance stats
elapsed_time = time.time() - start_time
self.retrieval_stats["total_retrievals"] += 1
self.retrieval_stats["total_time"] += elapsed_time
self.retrieval_stats["avg_time"] = (
self.retrieval_stats["total_time"] / self.retrieval_stats["total_retrievals"]
)
self.retrieval_stats["last_retrieval_time"] = elapsed_time
# Log performance summary
logger.info(f"β‘ RETRIEVAL PERFORMANCE: {elapsed_time*1000:.1f}ms total, {len(retrieval_results)}/{k} results")
logger.info(f"π RETRIEVAL COMPLETE: Query '{query}' processed successfully")
# Track performance using platform services
if self.platform:
self.platform.track_component_performance(
self,
"document_retrieval",
{
"success": True,
"retrieval_time": elapsed_time,
"query": query,
"results_count": len(retrieval_results),
"k_requested": k,
"indexed_documents": len(self.documents)
}
)
return retrieval_results
except Exception as e:
# Track failure using platform services
if self.platform:
elapsed_time = time.time() - start_time
self.platform.track_component_performance(
self,
"document_retrieval",
{
"success": False,
"retrieval_time": elapsed_time,
"query": query,
"k_requested": k,
"indexed_documents": len(self.documents),
"error": str(e)
}
)
logger.error(f"Modular retrieval failed: {str(e)}")
raise RuntimeError(f"Modular retrieval failed: {str(e)}") from e
def index_documents(self, documents: List[Document]) -> None:
"""
Index documents in all sub-components.
Args:
documents: List of documents to index
"""
if not documents:
raise ValueError("Cannot index empty document list")
# Store documents (extend existing instead of replacing)
if not hasattr(self, 'documents') or self.documents is None:
self.documents = []
self.documents.extend(documents)
# Get embedding dimension from first document
if documents[0].embedding is not None:
embedding_dim = len(documents[0].embedding)
else:
raise ValueError("Documents must have embeddings before indexing")
# Initialize index only if not already initialized
if not hasattr(self.vector_index, 'index') or self.vector_index.index is None:
self.vector_index.initialize_index(embedding_dim)
# Add documents to vector index
self.vector_index.add_documents(documents)
# Index in sparse retriever (this needs fixing too)
self.sparse_retriever.index_documents(documents)
logger.info(f"Indexed {len(documents)} documents in all sub-components")
def get_retrieval_stats(self) -> Dict[str, Any]:
"""
Get comprehensive statistics about the modular retrieval system.
Returns:
Dictionary with retrieval statistics and sub-component information
"""
stats = {
"component_type": "modular_unified_retriever",
"indexed_documents": len(self.documents),
"retrieval_stats": self.retrieval_stats.copy(),
"sub_components": {
"vector_index": self.vector_index.get_index_info(),
"sparse_retriever": self.sparse_retriever.get_stats(),
"fusion_strategy": self.fusion_strategy.get_strategy_info(),
"reranker": self.reranker.get_reranker_info()
}
}
return stats
def get_component_info(self) -> Dict[str, Any]:
"""
Get detailed information about all sub-components.
Returns:
Dictionary with component details for logging
"""
return {
"vector_index": self.vector_index.get_component_info(),
"sparse_retriever": self.sparse_retriever.get_component_info(),
"fusion_strategy": self.fusion_strategy.get_component_info(),
"reranker": self.reranker.get_component_info()
}
def supports_batch_queries(self) -> bool:
"""
Check if this retriever supports batch query processing.
Returns:
False, as the current implementation processes queries individually
"""
return False
def get_configuration(self) -> Dict[str, Any]:
"""
Get the current configuration of the modular retriever.
Returns:
Dictionary with configuration parameters
"""
return {
"vector_index": {
"type": "faiss",
"config": self.vector_index.get_index_info()
},
"sparse": {
"type": "bm25",
"config": self.sparse_retriever.get_stats()
},
"fusion": {
"type": type(self.fusion_strategy).__name__.lower().replace("fusion", ""),
"config": self.fusion_strategy.get_strategy_info()
},
"reranker": {
"type": type(self.reranker).__name__.lower().replace("reranker", ""),
"config": self.reranker.get_reranker_info()
}
}
def clear_index(self) -> None:
"""
Clear all indexed documents and reset all sub-components.
"""
self.documents.clear()
self.vector_index.clear()
self.sparse_retriever.clear()
# Reset performance stats
self.retrieval_stats = {
"total_retrievals": 0,
"total_time": 0.0,
"avg_time": 0.0,
"last_retrieval_time": 0.0
}
logger.info("Cleared all documents from modular retriever")
def _consider_backend_switch(self, error: Exception) -> None:
"""
Consider switching to a different backend due to an error.
This method is used for performance testing of backend switching.
In a real implementation, this would switch to a fallback backend.
Args:
error: The exception that triggered the switch consideration
"""
logger.warning(f"Backend switch consideration triggered by: {error}")
# Simulate backend switching logic
if self.active_backend_name == "faiss":
self.active_backend_name = "weaviate"
else:
self.active_backend_name = "faiss"
self.backend_switch_count += 1
logger.info(f"Switched to backend: {self.active_backend_name} (switch count: {self.backend_switch_count})")
# Standard ComponentBase interface implementation
def initialize_services(self, platform: 'PlatformOrchestrator') -> None:
"""Initialize platform services for the component.
Args:
platform: PlatformOrchestrator instance providing services
"""
self.platform = platform
logger.info("ModularUnifiedRetriever initialized with platform services")
def get_health_status(self) -> HealthStatus:
"""Get the current health status of the component.
Returns:
HealthStatus object with component health information
"""
if self.platform:
return self.platform.check_component_health(self)
# Fallback if platform services not initialized
is_healthy = True
issues = []
# Check sub-components
if not hasattr(self.vector_index, 'get_index_info'):
is_healthy = False
issues.append("Vector index not properly initialized")
if not hasattr(self.sparse_retriever, 'get_stats'):
is_healthy = False
issues.append("Sparse retriever not properly initialized")
if not hasattr(self.fusion_strategy, 'get_strategy_info'):
is_healthy = False
issues.append("Fusion strategy not properly initialized")
if not hasattr(self.reranker, 'get_reranker_info'):
is_healthy = False
issues.append("Reranker not properly initialized")
return HealthStatus(
is_healthy=is_healthy,
issues=issues,
metrics={
"indexed_documents": len(self.documents),
"retrieval_stats": self.retrieval_stats,
"sub_components": self.get_component_info()
},
component_name=self.__class__.__name__
)
def get_metrics(self) -> Dict[str, Any]:
"""Get component-specific metrics.
Returns:
Dictionary containing component metrics
"""
if self.platform:
try:
component_metrics = self.platform.analytics_service.collect_component_metrics(self)
return {
"component_name": component_metrics.component_name,
"component_type": component_metrics.component_type,
"success_count": component_metrics.success_count,
"error_count": component_metrics.error_count,
"resource_usage": component_metrics.resource_usage,
"performance_metrics": component_metrics.performance_metrics,
"timestamp": component_metrics.timestamp
}
except Exception as e:
# Fallback if platform service fails
pass
# Fallback if platform services not initialized
return {
"indexed_documents": len(self.documents),
"retrieval_stats": self.retrieval_stats,
"sub_components": self.get_component_info(),
"configuration": self.get_configuration()
}
def get_capabilities(self) -> List[str]:
"""Get list of component capabilities.
Returns:
List of capability strings
"""
capabilities = [
"hybrid_retrieval",
"modular_architecture",
"dense_search",
"sparse_search",
"result_fusion",
"reranking"
]
# Add vector index capabilities
if hasattr(self.vector_index, 'get_capabilities'):
capabilities.extend([f"vector_{cap}" for cap in self.vector_index.get_capabilities()])
# Add sparse retriever capabilities
if hasattr(self.sparse_retriever, 'get_capabilities'):
capabilities.extend([f"sparse_{cap}" for cap in self.sparse_retriever.get_capabilities()])
# Add fusion strategy capabilities
if hasattr(self.fusion_strategy, 'get_capabilities'):
capabilities.extend([f"fusion_{cap}" for cap in self.fusion_strategy.get_capabilities()])
# Add reranker capabilities
if hasattr(self.reranker, 'get_capabilities'):
capabilities.extend([f"reranker_{cap}" for cap in self.reranker.get_capabilities()])
return capabilities
def get_document_count(self) -> int:
"""Get the number of documents in the retriever."""
return len(self.documents)
def get_sub_component_performance(self) -> Dict[str, Any]:
"""
Get performance information for each sub-component.
Returns:
Dictionary with performance metrics
"""
performance = {
"vector_index": {
"document_count": self.vector_index.get_document_count(),
"is_trained": self.vector_index.is_trained()
},
"sparse_retriever": {
"document_count": self.sparse_retriever.get_document_count(),
"stats": self.sparse_retriever.get_stats()
},
"fusion_strategy": {
"info": self.fusion_strategy.get_strategy_info()
},
"reranker": {
"enabled": self.reranker.is_enabled(),
"info": self.reranker.get_reranker_info()
}
}
return performance
def debug_retrieval(self, query: str, k: int = 5) -> Dict[str, Any]:
"""
Perform retrieval with detailed debugging information.
Args:
query: Search query
k: Number of results to return
Returns:
Dictionary with step-by-step retrieval information
"""
debug_info = {
"query": query,
"k": k,
"steps": {}
}
try:
# Step 1: Query embedding
query_embedding = self.embedder.embed_query(query)
debug_info["steps"]["embedding"] = {
"embedding_dim": len(query_embedding),
"embedding_norm": float(np.linalg.norm(query_embedding))
}
# Step 2: Dense search
dense_results = self.vector_index.search(query_embedding, k=k*2)
debug_info["steps"]["dense_search"] = {
"results_count": len(dense_results),
"top_scores": [score for _, score in dense_results[:5]]
}
# Step 3: Sparse search
sparse_results = self.sparse_retriever.search(query, k=k*2)
debug_info["steps"]["sparse_search"] = {
"results_count": len(sparse_results),
"top_scores": [score for _, score in sparse_results[:5]]
}
# Step 3.5: Set documents and query for graph enhancement (if supported)
if hasattr(self.fusion_strategy, 'set_documents_and_query'):
self.fusion_strategy.set_documents_and_query(self.documents, query)
# Step 4: Fusion
fused_results = self.fusion_strategy.fuse_results(dense_results, sparse_results)
debug_info["steps"]["fusion"] = {
"results_count": len(fused_results),
"top_scores": [score for _, score in fused_results[:5]],
"fusion_type": type(self.fusion_strategy).__name__
}
# Step 5: Reranking
if self.reranker.is_enabled():
top_candidates = fused_results[:k*2]
candidate_documents = [self.documents[idx] for idx, _ in top_candidates]
candidate_scores = [score for _, score in top_candidates]
reranked_results = self.reranker.rerank(query, candidate_documents, candidate_scores)
debug_info["steps"]["reranking"] = {
"enabled": True,
"candidates_count": len(candidate_documents),
"reranked_count": len(reranked_results),
"top_reranked_scores": [score for _, score in reranked_results[:5]]
}
else:
debug_info["steps"]["reranking"] = {"enabled": False}
except Exception as e:
debug_info["error"] = str(e)
return debug_info
def _calculate_composite_scores(self, query_embedding: np.ndarray, fused_results: List[Tuple[int, float]]) -> List[Tuple[int, float]]:
"""
Calculate composite scores for individual documents combining fusion scores and semantic similarity.
This method replaces the global semantic gap detection with per-document quality assessment.
Each document gets a composite score: Ξ± * fusion_score + Ξ² * semantic_similarity
Only documents above the composite threshold are included.
Args:
query_embedding: Query embedding vector
fused_results: List of (document_index, fusion_score) from fusion strategy
Returns:
List of (document_index, composite_score) for documents that pass the threshold
"""
if not fused_results:
return []
try:
# Normalize fusion scores to 0-1 range for fair combination
fusion_scores = [score for _, score in fused_results]
if len(set(fusion_scores)) > 1: # Only normalize if there's variation
min_score, max_score = min(fusion_scores), max(fusion_scores)
score_range = max_score - min_score
if score_range > 0:
normalized_fusion = [(score - min_score) / score_range for score in fusion_scores]
else:
normalized_fusion = [1.0] * len(fusion_scores) # All scores identical
else:
normalized_fusion = [1.0] * len(fusion_scores) # All scores identical
# Get document texts and embeddings
doc_indices = [doc_idx for doc_idx, _ in fused_results]
documents = [self.documents[idx] for idx in doc_indices if idx < len(self.documents)]
if not documents:
return []
doc_texts = [doc.content for doc in documents]
doc_embeddings = self.embedder.embed(doc_texts)
# Calculate composite scores for each document
composite_results = []
for i, (doc_idx, original_fusion_score) in enumerate(fused_results):
if i >= len(doc_embeddings) or doc_idx >= len(self.documents):
continue
# Calculate semantic similarity
doc_emb_array = np.array(doc_embeddings[i])
query_norm = query_embedding / np.linalg.norm(query_embedding)
doc_norm = doc_emb_array / np.linalg.norm(doc_emb_array)
semantic_similarity = np.dot(query_norm, doc_norm)
# Calculate composite score
normalized_fusion_score = normalized_fusion[i]
composite_score = (self.fusion_weight * normalized_fusion_score +
self.semantic_weight * semantic_similarity)
# Apply threshold filter
if composite_score >= self.min_composite_score:
composite_results.append((doc_idx, composite_score))
# Debug logging for first few documents
if i < 3:
logger.info(f"COMPOSITE DEBUG - Doc {i+1}: fusion={original_fusion_score:.3f}, "
f"norm_fusion={normalized_fusion_score:.3f}, semantic={semantic_similarity:.3f}, "
f"composite={composite_score:.3f}, threshold={self.min_composite_score}")
# Sort by composite score (descending) and return
composite_results.sort(key=lambda x: x[1], reverse=True)
logger.info(f"COMPOSITE FILTERING - {len(fused_results)} input β {len(composite_results)} passed threshold")
return composite_results
except Exception as e:
logger.warning(f"Error in composite scoring: {e}")
# Fallback to original fusion results
return fused_results
def _calculate_semantic_alignment(self, query_embedding: np.ndarray, fused_results: List[Tuple[int, float]]) -> float:
"""
Calculate semantic alignment between query and top retrieved documents.
Args:
query_embedding: Query embedding vector
fused_results: List of (document_index, score) from fusion
Returns:
Average cosine similarity between query and top documents
"""
if not fused_results:
return 0.0
try:
# Get embeddings for top documents
top_doc_indices = [doc_idx for doc_idx, _ in fused_results]
top_documents = [self.documents[idx] for idx in top_doc_indices if idx < len(self.documents)]
if not top_documents:
return 0.0
# Extract text content for embedding
doc_texts = [doc.content for doc in top_documents]
# Get document embeddings
doc_embeddings = self.embedder.embed(doc_texts)
# Calculate cosine similarities
similarities = []
for i, doc_embedding in enumerate(doc_embeddings):
doc_emb_array = np.array(doc_embedding)
# Normalize vectors for cosine similarity
query_norm = query_embedding / np.linalg.norm(query_embedding)
doc_norm = doc_emb_array / np.linalg.norm(doc_emb_array)
similarity = np.dot(query_norm, doc_norm)
similarities.append(similarity)
# Debug: Log individual document similarities for investigation
if i < 3: # Only log first 3 docs to avoid spam
doc_preview = doc_texts[i][:100] + "..." if len(doc_texts[i]) > 100 else doc_texts[i]
logger.debug(f"Doc {i+1} similarity: {similarity:.3f} - {doc_preview}")
# Return average similarity
avg_similarity = np.mean(similarities) if similarities else 0.0
logger.debug(f"Semantic alignment: {len(similarities)} docs, similarities={[f'{s:.3f}' for s in similarities[:5]]}, avg={avg_similarity:.3f}")
return float(avg_similarity)
except Exception as e:
logger.warning(f"Error calculating semantic alignment: {e}")
return 0.0 # Conservative fallback |