Spaces:
Running
Running
File size: 11,161 Bytes
6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 bcbc40e 6f35377 | 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 | import os
import re
import uuid
from typing import List, Dict, Any, Optional
from qdrant_client import QdrantClient
from qdrant_client.models import (
Distance,
VectorParams,
PointStruct,
Filter,
FieldCondition,
MatchValue,
PayloadSchemaType
)
from sentence_transformers import SentenceTransformer, CrossEncoder
from rank_bm25 import BM25Okapi
from core.config import settings
from core.logger import logger
class VectorStore:
_instance = None
def __new__(cls):
if cls._instance is None:
cls._instance = super(VectorStore, cls).__new__(cls)
try:
cls._instance.client = QdrantClient(
url=settings.QDRANT_URL,
api_key=settings.QDRANT_API_KEY,
timeout=5.0
)
cls._instance.client.get_collections()
logger.info(f"Connected to online Qdrant cloud database at {settings.QDRANT_URL}")
except Exception as e:
logger.warning(
f"Remote Qdrant Cloud cluster unreachable ({e}). "
f"Falling back to local in-memory Qdrant instance."
)
cls._instance.client = QdrantClient(location=":memory:")
cls._instance.collection_name = "ecommerce_chunks"
cls._instance._ensure_collection()
logger.info(f"Initializing SentenceTransformer embedding model on CPU ({settings.EMBEDDING_MODEL_NAME})")
cls._instance.embedding_model = SentenceTransformer(settings.EMBEDDING_MODEL_NAME)
cls._instance.embedding_model.to('cpu')
logger.info(f"Initializing CrossEncoder reranker model on CPU ({settings.RERANKER_MODEL_NAME})")
cls._instance.reranker = CrossEncoder(settings.RERANKER_MODEL_NAME)
return cls._instance
def _ensure_collection(self):
try:
collections = self.client.get_collections().collections
exists = any(c.name == self.collection_name for c in collections)
if not exists:
logger.info(f"Creating Qdrant collection: {self.collection_name}")
self.client.create_collection(
collection_name=self.collection_name,
vectors_config=VectorParams(size=384, distance=Distance.COSINE)
)
# Ensure payload index on session_id for fast isolated filtering
try:
self.client.create_payload_index(
collection_name=self.collection_name,
field_name="session_id",
field_schema=PayloadSchemaType.KEYWORD
)
except Exception:
pass # Index might already exist
except Exception as e:
logger.error(f"Error ensuring Qdrant collection: {str(e)}")
raise
def add_chunks(self, chunks: List[Dict[str, Any]], session_id: Optional[str] = None):
"""
Add chunks to Qdrant collection with session isolation and parent-document linkage.
"""
if not chunks:
return
active_session = session_id or "global"
logger.info(f"Generating embeddings and indexing {len(chunks)} chunks in Qdrant (Session: {active_session})")
texts = [chunk["document"] for chunk in chunks]
# Generate dense embeddings
embeddings = self.embedding_model.encode(texts, convert_to_numpy=True, show_progress_bar=False)
points = []
for i, chunk in enumerate(chunks):
point_id = chunk["id"]
if len(point_id) < 36:
point_id = str(uuid.uuid5(uuid.NAMESPACE_DNS, point_id))
chunk_session = chunk.get("session_id") or active_session
payload = {
"document": chunk["document"],
"raw_text": chunk.get("raw_text", chunk["document"]),
"source": chunk.get("source", ""),
"chunk_type": chunk.get("chunk_type", "child"),
"created_at": chunk.get("created_at", ""),
"scraped_timestamp": chunk.get("scraped_timestamp", ""),
"sequence": chunk.get("sequence", None),
"parent_id": chunk.get("parent_id", None),
"parent_content": chunk.get("parent_content", chunk["document"]),
"session_id": chunk_session,
"structured_data": chunk.get("structured_data", {})
}
points.append(PointStruct(
id=point_id,
vector=embeddings[i].tolist(),
payload=payload
))
self.client.upsert(
collection_name=self.collection_name,
points=points
)
logger.info(f"Successfully indexed {len(points)} chunks in Qdrant for session '{active_session}'")
def search(
self,
query_text: str,
k: int = 5,
session_id: Optional[str] = None
) -> List[Dict[str, Any]]:
"""
Hybrid retrieval with dense vectors, BM25, Cross-Encoder reranking, and parent document resolution.
"""
logger.info(f"Hybrid SOTA Search for: '{query_text[:50]}' (Session: {session_id or 'all'})")
try:
query_vector = self.embedding_model.encode(
query_text,
convert_to_numpy=True,
show_progress_bar=False
).tolist()
candidate_pool_size = max(k * 4, 15)
query_filter = None
if session_id and session_id != "global":
query_filter = Filter(must=[
FieldCondition(key="session_id", match=MatchValue(value=session_id))
])
response = self.client.query_points(
collection_name=self.collection_name,
query=query_vector,
query_filter=query_filter,
limit=candidate_pool_size
)
hits = response.points
if not hits and query_filter is not None:
logger.info("No session-specific chunks found; falling back to broader corpus")
response = self.client.query_points(
collection_name=self.collection_name,
query=query_vector,
limit=candidate_pool_size
)
hits = response.points
if not hits:
return []
candidates = []
for hit in hits:
payload = hit.payload or {}
candidates.append({
"id": hit.id,
"document": payload.get("document", ""),
"raw_text": payload.get("raw_text", payload.get("document", "")),
"parent_content": payload.get("parent_content", payload.get("document", "")),
"parent_id": payload.get("parent_id") or hit.id,
"metadata": {
"source": payload.get("source", ""),
"chunk_type": payload.get("chunk_type", ""),
"created_at": payload.get("created_at", ""),
"scraped_timestamp": payload.get("scraped_timestamp", ""),
"sequence": payload.get("sequence", None),
"parent_id": payload.get("parent_id", None),
"session_id": payload.get("session_id", ""),
"structured_data": payload.get("structured_data", {})
},
"dense_score": float(hit.score)
})
tokenized_corpus = [
re.findall(r'\w+', (c["document"] + " " + str(c["metadata"].get("structured_data", ""))).lower())
for c in candidates
]
tokenized_query = re.findall(r'\w+', query_text.lower())
if tokenized_corpus and tokenized_query:
bm25 = BM25Okapi(tokenized_corpus)
bm25_scores = bm25.get_scores(tokenized_query)
bm25_ranked_indices = sorted(range(len(candidates)), key=lambda i: bm25_scores[i], reverse=True)
bm25_ranks = {candidates[idx]["id"]: rank for rank, idx in enumerate(bm25_ranked_indices)}
dense_ranked_indices = sorted(range(len(candidates)), key=lambda i: candidates[i]["dense_score"], reverse=True)
dense_ranks = {candidates[idx]["id"]: rank for rank, idx in enumerate(dense_ranked_indices)}
for c in candidates:
cid = c["id"]
d_rank = dense_ranks.get(cid, len(candidates))
s_rank = bm25_ranks.get(cid, len(candidates))
c["rrf_score"] = (1.0 / (60.0 + d_rank)) + (1.0 / (60.0 + s_rank))
candidates.sort(key=lambda x: x.get("rrf_score", 0.0), reverse=True)
logger.info(f"Applied BM25 + Dense RRF fusion over {len(candidates)} candidates")
if self.reranker:
try:
pairs = [(query_text, c["document"][:800]) for c in candidates]
rerank_scores = self.reranker.predict(pairs)
for i, score in enumerate(rerank_scores):
candidates[i]["rerank_score"] = float(score)
candidates.sort(key=lambda x: x.get("rerank_score", -999.0), reverse=True)
logger.info("Successfully re-ranked candidates using Cross-Encoder")
except Exception as e:
logger.error(f"Cross-encoder reranking error: {e}")
seen_parent_ids = set()
parent_results = []
for c in candidates:
parent_id = c["parent_id"]
if parent_id in seen_parent_ids:
continue
seen_parent_ids.add(parent_id)
full_parent_doc = c["parent_content"] or c["document"]
parent_results.append({
"id": c["id"],
"parent_id": parent_id,
"document": full_parent_doc,
"child_match_snippet": c["document"][:300],
"metadata": c["metadata"],
"score": c.get("rerank_score", c.get("rrf_score", c["dense_score"]))
})
if len(parent_results) >= k:
break
logger.info(f"Retrieved and resolved {len(parent_results)} parent documents after SOTA re-ranking")
return parent_results
except Exception as e:
logger.error(f"Qdrant SOTA search error: {str(e)}")
return []
|