File size: 9,303 Bytes
9b457ed | 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 | """
Embedding generation supporting multiple model backends.
This module provides efficient batch embedding generation with automatic
model loading, caching, and device management. Supports both SentenceTransformers
models and NVIDIA NV-Embed-v2.
"""
import numpy as np
import torch
from typing import List, Optional
from tqdm import tqdm
from src.config.settings import get_settings, get_embedding_model_config, EMBEDDING_MODELS
from src.utils.logging import get_logger, log_embedding_generation
from src.ingestion.models import Chunk
import time
logger = get_logger(__name__)
class Embedder:
"""Generate embeddings using SentenceTransformers or NV-Embed-v2."""
# Models that require special handling
NVEMBED_MODELS = ["nvidia/NV-Embed-v2", "nvidia/NV-Embed-v1"]
def __init__(self, model_name: Optional[str] = None):
"""
Initialize embedder with specified or default model.
Args:
model_name: Optional model identifier. If None, uses settings default.
"""
settings = get_settings()
self.model_name = model_name or settings.embedding_model
self.device = settings.embedding_device
# Get model-specific config
try:
model_config = get_embedding_model_config(self.model_name)
self.batch_size = model_config.get("batch_size", settings.embedding_batch_size)
self._dimensions = model_config.get("dimensions")
self._max_length = model_config.get("max_length", 512)
except ValueError:
# Fallback for unknown models
self.batch_size = settings.embedding_batch_size
self._dimensions = None
self._max_length = 512
self._model = None
self._tokenizer = None
self._is_nvembed = self.model_name in self.NVEMBED_MODELS
@property
def model(self):
"""
Lazy load the embedding model.
The model is only loaded when first accessed, and then cached for reuse.
Returns:
Model instance (SentenceTransformer or transformers model)
"""
if self._model is None:
logger.info(f"Loading embedding model: {self.model_name}")
if self._is_nvembed:
self._load_nvembed_model()
else:
self._load_sentence_transformer()
logger.info(f"Model loaded on device: {self.device}")
return self._model
def _load_sentence_transformer(self):
"""Load a SentenceTransformer model."""
from sentence_transformers import SentenceTransformer
self._model = SentenceTransformer(self.model_name)
self._model.to(self.device)
def _load_nvembed_model(self):
"""Load NVIDIA NV-Embed-v2 model."""
from transformers import AutoModel, AutoTokenizer
logger.info("Loading NV-Embed-v2 (this may take a moment)...")
# Determine torch dtype based on device
if self.device == "mps":
# MPS works best with float32 for this model
torch_dtype = torch.float32
elif self.device == "cuda":
torch_dtype = torch.float16
else:
torch_dtype = torch.float32
self._tokenizer = AutoTokenizer.from_pretrained(
self.model_name,
trust_remote_code=True
)
self._model = AutoModel.from_pretrained(
self.model_name,
trust_remote_code=True,
torch_dtype=torch_dtype,
)
self._model.to(self.device)
self._model.eval()
def _nvembed_encode(
self,
texts: List[str],
instruction: str = "",
max_length: Optional[int] = None,
) -> np.ndarray:
"""
Encode texts using NV-Embed-v2's native encode method.
Args:
texts: List of texts to encode
instruction: Instruction prefix for queries (empty for documents)
max_length: Maximum sequence length (uses model config if None)
Returns:
np.ndarray: Embeddings array
"""
if max_length is None:
max_length = self._max_length
all_embeddings = []
for i in tqdm(range(0, len(texts), self.batch_size), desc="Embedding"):
batch_texts = texts[i:i + self.batch_size]
# Use NV-Embed's native encode method
with torch.no_grad():
if instruction:
# For queries: use instruction
embeddings = self._model.encode(
batch_texts,
instruction=instruction,
max_length=max_length,
)
else:
# For documents: no instruction needed
embeddings = self._model.encode(
batch_texts,
max_length=max_length,
)
# Handle both tensor and numpy outputs
if isinstance(embeddings, torch.Tensor):
embeddings = embeddings.cpu().numpy()
all_embeddings.append(embeddings)
return np.vstack(all_embeddings)
def encode_batch(self, chunks: List[Chunk]) -> np.ndarray:
"""
Generate embeddings for a batch of chunks (documents).
Processes chunks in smaller batches for memory efficiency and
displays progress with tqdm.
Args:
chunks: List of chunks to embed
Returns:
np.ndarray: Array of embeddings with shape (num_chunks, embedding_dim)
"""
if not chunks:
logger.warning("No chunks to embed")
return np.array([])
start_time = time.time()
# Extract text from chunks
texts = [chunk.text for chunk in chunks]
logger.info(f"Generating embeddings for {len(chunks)} chunks")
if self._is_nvembed:
# NV-Embed: documents don't need instruction prefix
_ = self.model # Ensure model is loaded
embeddings = self._nvembed_encode(texts, instruction="")
else:
# SentenceTransformers path
embeddings = []
for i in tqdm(range(0, len(texts), self.batch_size), desc="Embedding chunks"):
batch_texts = texts[i:i + self.batch_size]
batch_embeddings = self.model.encode(
batch_texts,
batch_size=self.batch_size,
show_progress_bar=False,
convert_to_numpy=True,
normalize_embeddings=True
)
embeddings.append(batch_embeddings)
embeddings = np.vstack(embeddings)
# Log performance
duration = time.time() - start_time
log_embedding_generation(logger, len(chunks), duration)
return embeddings
def encode_single(self, text: str, is_query: bool = False) -> np.ndarray:
"""
Generate embedding for a single text.
Args:
text: Text to embed
is_query: If True, applies query instruction (for NV-Embed)
Returns:
np.ndarray: Embedding vector
"""
if self._is_nvembed:
_ = self.model # Ensure model is loaded
# NV-Embed uses instruction prefix for queries
instruction = (
"Instruct: Given a question, retrieve passages that answer the question\nQuery: "
if is_query else ""
)
embeddings = self._nvembed_encode([text], instruction=instruction)
return embeddings[0]
else:
embedding = self.model.encode(
text,
convert_to_numpy=True,
normalize_embeddings=True
)
return embedding
def get_embedding_dimension(self) -> int:
"""
Get the dimension of embeddings produced by this model.
Returns:
int: Embedding dimension
"""
# Use pre-configured dimensions if available
if self._dimensions:
return self._dimensions
# Otherwise, load model and query
_ = self.model # Ensure model is loaded
if self._is_nvembed:
return 4096
else:
return self._model.get_sentence_embedding_dimension()
def get_model_info(self) -> dict:
"""
Get information about the current embedding model.
Returns:
dict: Model information including name, dimensions, etc.
"""
try:
config = get_embedding_model_config(self.model_name)
return {
"id": self.model_name,
"name": config.get("name", self.model_name),
"dimensions": self.get_embedding_dimension(),
"type": config.get("type", "unknown"),
"description": config.get("description", ""),
}
except ValueError:
return {
"id": self.model_name,
"name": self.model_name.split("/")[-1],
"dimensions": self.get_embedding_dimension(),
"type": "unknown",
"description": "Custom model",
}
|