Spaces:
Running
Running
File size: 9,987 Bytes
b5246f1 |
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 |
"""
Hybrid retrieval combining dense semantic search with sparse BM25 keyword matching.
Uses Reciprocal Rank Fusion (RRF) to combine results from both approaches.
"""
from typing import List, Dict, Tuple, Optional
import numpy as np
from pathlib import Path
import sys
# Add project root to Python path for imports
project_root = Path(__file__).parent.parent.parent / "project-1-technical-rag"
sys.path.append(str(project_root))
from src.sparse_retrieval import BM25SparseRetriever
from src.fusion import reciprocal_rank_fusion, adaptive_fusion
from shared_utils.embeddings.generator import generate_embeddings
import faiss
class HybridRetriever:
"""
Hybrid retrieval system combining dense semantic search with sparse BM25.
Optimized for technical documentation where both semantic similarity
and exact keyword matching are important for retrieval quality.
Performance: Sub-second search on 1000+ document corpus
"""
def __init__(
self,
dense_weight: float = 0.7,
embedding_model: str = "sentence-transformers/multi-qa-MiniLM-L6-cos-v1",
use_mps: bool = True,
bm25_k1: float = 1.2,
bm25_b: float = 0.75,
rrf_k: int = 10
):
"""
Initialize hybrid retriever with dense and sparse components.
Args:
dense_weight: Weight for semantic similarity in fusion (0.7 default)
embedding_model: Sentence transformer model name
use_mps: Use Apple Silicon MPS acceleration for embeddings
bm25_k1: BM25 term frequency saturation parameter
bm25_b: BM25 document length normalization parameter
rrf_k: Reciprocal Rank Fusion constant (1=strong rank preference, 2=moderate)
Raises:
ValueError: If parameters are invalid
"""
if not 0 <= dense_weight <= 1:
raise ValueError("dense_weight must be between 0 and 1")
self.dense_weight = dense_weight
self.embedding_model = embedding_model
self.use_mps = use_mps
self.rrf_k = rrf_k
# Initialize sparse retriever
self.sparse_retriever = BM25SparseRetriever(k1=bm25_k1, b=bm25_b)
# Dense retrieval components (initialized on first index)
self.dense_index: Optional[faiss.Index] = None
self.chunks: List[Dict] = []
self.embeddings: Optional[np.ndarray] = None
def index_documents(self, chunks: List[Dict]) -> None:
"""
Index documents for both dense and sparse retrieval.
Args:
chunks: List of chunk dictionaries with 'text' field
Raises:
ValueError: If chunks is empty or malformed
Performance: ~100 chunks/second for complete indexing
"""
if not chunks:
raise ValueError("Cannot index empty chunk list")
print(f"Indexing {len(chunks)} chunks for hybrid retrieval...")
# Store chunks for retrieval
self.chunks = chunks
# Index for sparse retrieval
print("Building BM25 sparse index...")
self.sparse_retriever.index_documents(chunks)
# Index for dense retrieval
print("Building dense semantic index...")
texts = [chunk['text'] for chunk in chunks]
# Generate embeddings
self.embeddings = generate_embeddings(
texts,
model_name=self.embedding_model,
use_mps=self.use_mps
)
# Create FAISS index
embedding_dim = self.embeddings.shape[1]
self.dense_index = faiss.IndexFlatIP(embedding_dim) # Inner product for cosine similarity
# Normalize embeddings for cosine similarity
faiss.normalize_L2(self.embeddings)
self.dense_index.add(self.embeddings)
print(f"Hybrid indexing complete: {len(chunks)} chunks ready for search")
def search(
self,
query: str,
top_k: int = 10,
dense_top_k: Optional[int] = None,
sparse_top_k: Optional[int] = None
) -> List[Tuple[int, float, Dict]]:
"""
Hybrid search combining dense and sparse retrieval with RRF.
Args:
query: Search query string
top_k: Final number of results to return
dense_top_k: Results from dense search (default: 2*top_k)
sparse_top_k: Results from sparse search (default: 2*top_k)
Returns:
List of (chunk_index, rrf_score, chunk_dict) tuples
Raises:
ValueError: If not indexed or invalid parameters
Performance: <200ms for 1000+ document corpus
"""
if self.dense_index is None:
raise ValueError("Must call index_documents() before searching")
if not query.strip():
return []
if top_k <= 0:
raise ValueError("top_k must be positive")
# Set default intermediate result counts
if dense_top_k is None:
dense_top_k = min(2 * top_k, len(self.chunks))
if sparse_top_k is None:
sparse_top_k = min(2 * top_k, len(self.chunks))
# Dense semantic search
dense_results = self._dense_search(query, dense_top_k)
# Sparse BM25 search
sparse_results = self.sparse_retriever.search(query, sparse_top_k)
# Combine using Adaptive Fusion (better for small result sets)
fused_results = adaptive_fusion(
dense_results=dense_results,
sparse_results=sparse_results,
dense_weight=self.dense_weight,
result_size=top_k
)
# Prepare final results with chunk content and apply source diversity
final_results = []
for chunk_idx, rrf_score in fused_results:
chunk_dict = self.chunks[chunk_idx]
final_results.append((chunk_idx, rrf_score, chunk_dict))
# Apply source diversity enhancement
diverse_results = self._enhance_source_diversity(final_results, top_k)
return diverse_results
def _dense_search(self, query: str, top_k: int) -> List[Tuple[int, float]]:
"""
Perform dense semantic search using FAISS.
Args:
query: Search query
top_k: Number of results to return
Returns:
List of (chunk_index, similarity_score) tuples
"""
# Generate query embedding
query_embedding = generate_embeddings(
[query],
model_name=self.embedding_model,
use_mps=self.use_mps
)
# Normalize for cosine similarity
faiss.normalize_L2(query_embedding)
# Search dense index
similarities, indices = self.dense_index.search(query_embedding, top_k)
# Convert to required format
results = [
(int(indices[0][i]), float(similarities[0][i]))
for i in range(len(indices[0]))
if indices[0][i] != -1 # Filter out invalid results
]
return results
def _enhance_source_diversity(
self,
results: List[Tuple[int, float, Dict]],
top_k: int,
max_per_source: int = 2
) -> List[Tuple[int, float, Dict]]:
"""
Enhance source diversity in retrieval results to prevent over-focusing on single documents.
Args:
results: List of (chunk_idx, score, chunk_dict) tuples sorted by relevance
top_k: Maximum number of results to return
max_per_source: Maximum chunks allowed per source document
Returns:
Diversified results maintaining relevance while improving source coverage
"""
if not results:
return []
source_counts = {}
diverse_results = []
# First pass: Add highest scoring results respecting source limits
for chunk_idx, score, chunk_dict in results:
source = chunk_dict.get('source', 'unknown')
current_count = source_counts.get(source, 0)
if current_count < max_per_source:
diverse_results.append((chunk_idx, score, chunk_dict))
source_counts[source] = current_count + 1
if len(diverse_results) >= top_k:
break
# Second pass: If we still need more results, relax source constraints
if len(diverse_results) < top_k:
for chunk_idx, score, chunk_dict in results:
if (chunk_idx, score, chunk_dict) not in diverse_results:
diverse_results.append((chunk_idx, score, chunk_dict))
if len(diverse_results) >= top_k:
break
return diverse_results[:top_k]
def get_retrieval_stats(self) -> Dict[str, any]:
"""
Get statistics about the indexed corpus and retrieval performance.
Returns:
Dictionary with corpus statistics
"""
if not self.chunks:
return {"status": "not_indexed"}
return {
"status": "indexed",
"total_chunks": len(self.chunks),
"dense_index_size": self.dense_index.ntotal if self.dense_index else 0,
"embedding_dim": self.embeddings.shape[1] if self.embeddings is not None else 0,
"sparse_indexed_chunks": len(self.sparse_retriever.chunk_mapping),
"dense_weight": self.dense_weight,
"sparse_weight": 1.0 - self.dense_weight,
"rrf_k": self.rrf_k
} |