File size: 14,615 Bytes
c5880fb |
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 |
"""
Embedding Service Module
Provides multilingual semantic search using sentence-transformers.
Uses paraphrase-multilingual-MiniLM-L12-v2 by default which supports 50+ languages
including English, Ukrainian, Russian, Spanish, German, French, etc.
References:
- Reimers & Gurevych (2019): Sentence-BERT
- Reimers & Gurevych (2020): Making Monolingual Sentence Embeddings Multilingual
"""
import os
import json
import logging
import numpy as np
from pathlib import Path
from typing import Dict, List, Optional, Tuple, Any
from dataclasses import dataclass
import hashlib
logger = logging.getLogger(__name__)
@dataclass
class SearchResult:
"""Result from semantic search."""
entity_id: str
score: float
entity_data: Dict[str, Any]
class EmbeddingService:
"""
Multilingual embedding service for semantic search.
Replaces keyword-based matching with embedding similarity,
enabling language-agnostic symptom/entity matching.
"""
def __init__(
self,
model_name: str = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
cache_dir: str = "./data/embeddings",
device: str = "cpu"
):
self.model_name = model_name
self.cache_dir = Path(cache_dir)
self.cache_dir.mkdir(parents=True, exist_ok=True)
self.device = device
self._model = None
self._entity_embeddings: Dict[str, np.ndarray] = {}
self._entity_data: Dict[str, Dict] = {}
self._embedding_dim: int = 384
# Index for fast similarity search
self._index = None
self._index_ids: List[str] = []
@property
def model(self):
"""Lazy load the embedding model."""
if self._model is None:
try:
from sentence_transformers import SentenceTransformer
logger.info(f"Loading embedding model: {self.model_name}")
self._model = SentenceTransformer(self.model_name, device=self.device)
self._embedding_dim = self._model.get_sentence_embedding_dimension()
logger.info(f"Model loaded. Embedding dimension: {self._embedding_dim}")
except ImportError:
logger.error(
"sentence-transformers not installed. "
"Run: pip install sentence-transformers"
)
raise
return self._model
def encode(self, texts: List[str], batch_size: int = 32) -> np.ndarray:
"""
Encode texts to embeddings.
Args:
texts: List of text strings to encode
batch_size: Batch size for encoding
Returns:
numpy array of shape (len(texts), embedding_dim)
"""
if not texts:
return np.array([])
embeddings = self.model.encode(
texts,
batch_size=batch_size,
show_progress_bar=len(texts) > 100,
convert_to_numpy=True,
normalize_embeddings=True # For cosine similarity via dot product
)
return embeddings
def encode_single(self, text: str) -> np.ndarray:
"""Encode a single text string."""
return self.encode([text])[0]
def index_entities(
self,
entities: Dict[str, Dict[str, Any]],
text_fields: List[str] = ["name", "description", "synonyms"]
):
"""
Build search index from entities.
Args:
entities: Dict of entity_id -> entity_data
text_fields: Fields to combine for embedding text
"""
logger.info(f"Indexing {len(entities)} entities for semantic search")
# Check cache
cache_key = self._compute_cache_key(entities)
if self._load_from_cache(cache_key):
logger.info("Loaded embeddings from cache")
return
# Prepare texts for embedding
texts = []
entity_ids = []
for entity_id, entity in entities.items():
# Combine relevant text fields
text_parts = []
for field in text_fields:
value = entity.get(field)
if value:
if isinstance(value, list):
text_parts.extend(value)
else:
text_parts.append(str(value))
if text_parts:
combined_text = " ".join(text_parts)
texts.append(combined_text)
entity_ids.append(entity_id)
self._entity_data[entity_id] = entity
if not texts:
logger.warning("No texts to index")
return
# Compute embeddings
logger.info(f"Computing embeddings for {len(texts)} entities...")
embeddings = self.encode(texts)
# Store embeddings
for entity_id, embedding in zip(entity_ids, embeddings):
self._entity_embeddings[entity_id] = embedding
self._index_ids = entity_ids
# Build FAISS index if available, else use numpy
self._build_index(embeddings)
# Save to cache
self._save_to_cache(cache_key)
logger.info(f"Indexed {len(self._entity_embeddings)} entities")
def _build_index(self, embeddings: np.ndarray):
"""Build search index from embeddings."""
try:
import faiss
# Use IndexFlatIP for inner product (cosine similarity with normalized vectors)
self._index = faiss.IndexFlatIP(self._embedding_dim)
self._index.add(embeddings.astype(np.float32))
logger.info("Built FAISS index for fast similarity search")
except ImportError:
# Fallback to numpy-based search
logger.info("FAISS not available, using numpy for similarity search")
self._index = None
self._embedding_matrix = embeddings
def search(
self,
query: str,
top_k: int = 10,
threshold: float = 0.3,
category_filter: Optional[str] = None
) -> List[SearchResult]:
"""
Search for entities similar to query.
Args:
query: Search query (any language)
top_k: Maximum number of results
threshold: Minimum similarity score (0-1)
category_filter: Optional category to filter by
Returns:
List of SearchResult sorted by score descending
"""
if not self._entity_embeddings:
logger.warning("No entities indexed. Call index_entities first.")
return []
# Encode query
query_embedding = self.encode_single(query)
# Search
if self._index is not None:
# FAISS search
scores, indices = self._index.search(
query_embedding.reshape(1, -1).astype(np.float32),
min(top_k * 2, len(self._index_ids)) # Get more for filtering
)
scores = scores[0]
indices = indices[0]
else:
# Numpy fallback
scores = np.dot(self._embedding_matrix, query_embedding)
indices = np.argsort(scores)[::-1][:top_k * 2]
scores = scores[indices]
# Build results with filtering
results = []
for score, idx in zip(scores, indices):
if score < threshold:
continue
if idx < 0 or idx >= len(self._index_ids):
continue
entity_id = self._index_ids[idx]
entity_data = self._entity_data.get(entity_id, {})
# Apply category filter
if category_filter and entity_data.get("category") != category_filter:
continue
results.append(SearchResult(
entity_id=entity_id,
score=float(score),
entity_data=entity_data
))
if len(results) >= top_k:
break
return results
def search_multiple(
self,
queries: List[str],
top_k_per_query: int = 5,
threshold: float = 0.3,
deduplicate: bool = True
) -> List[SearchResult]:
"""
Search with multiple queries, combining results.
Useful for extracting multiple symptoms from a single user query.
"""
all_results: Dict[str, SearchResult] = {}
for query in queries:
results = self.search(query, top_k=top_k_per_query, threshold=threshold)
for result in results:
if result.entity_id not in all_results:
all_results[result.entity_id] = result
else:
# Keep highest score
if result.score > all_results[result.entity_id].score:
all_results[result.entity_id] = result
# Sort by score
return sorted(all_results.values(), key=lambda x: x.score, reverse=True)
def extract_entities_from_text(
self,
text: str,
category: Optional[str] = None,
top_k: int = 5,
threshold: float = 0.4
) -> List[SearchResult]:
"""
Extract relevant entities from free-form text.
This is the main method for symptom extraction from user queries.
Works across all supported languages.
Args:
text: User input text (any language)
category: Filter by category (e.g., "symptom", "disease")
top_k: Maximum entities to return
threshold: Minimum similarity threshold
"""
# Direct search on full text
results = self.search(
query=text,
top_k=top_k,
threshold=threshold,
category_filter=category
)
# Also try splitting into phrases (helps with multiple symptoms)
# Split on common separators
import re
phrases = re.split(r'[,;.]|\band\b|\bwith\b|\balso\b|\bі\b|\bта\b|\bи\b', text)
phrases = [p.strip() for p in phrases if p.strip() and len(p.strip()) > 2]
if len(phrases) > 1:
phrase_results = self.search_multiple(
phrases,
top_k_per_query=3,
threshold=threshold
)
# Merge results
seen_ids = {r.entity_id for r in results}
for pr in phrase_results:
if pr.entity_id not in seen_ids:
results.append(pr)
seen_ids.add(pr.entity_id)
# Sort and limit
results.sort(key=lambda x: x.score, reverse=True)
return results[:top_k]
def _compute_cache_key(self, entities: Dict) -> str:
"""Compute cache key from entities."""
# Hash based on entity IDs and model name
entity_str = json.dumps(sorted(entities.keys()))
key_str = f"{self.model_name}:{entity_str}"
return hashlib.md5(key_str.encode()).hexdigest()[:16]
def _load_from_cache(self, cache_key: str) -> bool:
"""Try to load embeddings from cache."""
embeddings_path = self.cache_dir / f"{cache_key}_embeddings.npy"
metadata_path = self.cache_dir / f"{cache_key}_metadata.json"
if not embeddings_path.exists() or not metadata_path.exists():
return False
try:
# Load metadata
with open(metadata_path) as f:
metadata = json.load(f)
# Verify model matches
if metadata.get("model") != self.model_name:
return False
# Load embeddings
embeddings = np.load(embeddings_path)
# Restore state
self._index_ids = metadata["entity_ids"]
self._entity_data = metadata["entity_data"]
for i, entity_id in enumerate(self._index_ids):
self._entity_embeddings[entity_id] = embeddings[i]
# Rebuild index
self._build_index(embeddings)
return True
except Exception as e:
logger.warning(f"Failed to load from cache: {e}")
return False
def _save_to_cache(self, cache_key: str):
"""Save embeddings to cache."""
try:
embeddings_path = self.cache_dir / f"{cache_key}_embeddings.npy"
metadata_path = self.cache_dir / f"{cache_key}_metadata.json"
# Prepare embeddings array
embeddings = np.array([
self._entity_embeddings[eid] for eid in self._index_ids
])
# Save embeddings
np.save(embeddings_path, embeddings)
# Save metadata
metadata = {
"model": self.model_name,
"entity_ids": self._index_ids,
"entity_data": self._entity_data,
"embedding_dim": self._embedding_dim
}
with open(metadata_path, "w") as f:
json.dump(metadata, f)
logger.info(f"Saved embeddings cache: {cache_key}")
except Exception as e:
logger.warning(f"Failed to save cache: {e}")
def clear_cache(self):
"""Clear all cached embeddings."""
import shutil
if self.cache_dir.exists():
shutil.rmtree(self.cache_dir)
self.cache_dir.mkdir(parents=True, exist_ok=True)
self._entity_embeddings.clear()
self._entity_data.clear()
self._index = None
self._index_ids = []
# Global instance (lazy initialized)
_embedding_service: Optional[EmbeddingService] = None
def get_embedding_service() -> EmbeddingService:
"""Get the global embedding service instance."""
global _embedding_service
if _embedding_service is None:
from .config import get_config
config = get_config()
_embedding_service = EmbeddingService(
model_name=config.embedding.model_name,
cache_dir=config.embedding.cache_dir,
device=config.embedding.device
)
return _embedding_service
|