Files
LangBot/pkg/rag/knowledge/services/embedding_models.py
WangCham 4bcc06c955 kb
2025-07-05 18:58:16 +08:00

223 lines
10 KiB
Python

# services/embedding_models.py
import os
from typing import Dict, Any, List, Type, Optional
import logging
import aiohttp # Import aiohttp for asynchronous requests
import asyncio
from sentence_transformers import SentenceTransformer
logger = logging.getLogger(__name__)
# Base class for all embedding models
class BaseEmbeddingModel:
def __init__(self, model_name: str):
self.model_name = model_name
self._embedding_dimension = None
async def embed_documents(self, texts: List[str]) -> List[List[float]]:
"""Asynchronously embeds a list of texts."""
raise NotImplementedError
async def embed_query(self, text: str) -> List[float]:
"""Asynchronously embeds a single query text."""
raise NotImplementedError
@property
def embedding_dimension(self) -> int:
"""Returns the embedding dimension of the model."""
if self._embedding_dimension is None:
raise NotImplementedError("Embedding dimension not set for this model.")
return self._embedding_dimension
class EmbeddingModelFactory:
@staticmethod
def create_model(model_type: str, model_name_key: str) -> BaseEmbeddingModel:
"""
Factory method to create an embedding model instance.
Currently only supports 'third_party_api' types.
"""
if model_name_key not in EMBEDDING_MODEL_CONFIGS:
raise ValueError(f"Embedding model configuration '{model_name_key}' not found in EMBEDDING_MODEL_CONFIGS.")
config = EMBEDDING_MODEL_CONFIGS[model_name_key]
if config['type'] == "third_party_api":
required_keys = ['api_endpoint', 'headers', 'payload_template', 'embedding_dimension']
if not all(key in config for key in required_keys):
raise ValueError(f"Missing configuration keys for third_party_api model '{model_name_key}'. Required: {required_keys}")
# Retrieve model_name from config if it differs from model_name_key
# Some APIs expect a specific 'model' value in the payload that might be different from the key
api_model_name = config.get('model_name', model_name_key)
return ThirdPartyAPIEmbeddingModel(
model_name=api_model_name, # Use the model_name from config or the key
api_endpoint=config['api_endpoint'],
headers=config['headers'],
payload_template=config['payload_template'],
embedding_dimension=config['embedding_dimension']
)
class SentenceTransformerEmbeddingModel(BaseEmbeddingModel):
def __init__(self, model_name: str):
super().__init__(model_name)
try:
# SentenceTransformer is inherently synchronous, but we'll wrap its calls
# in async methods. The actual computation will still block the event loop
# if not run in a separate thread/process, but this keeps the API consistent.
self.model = SentenceTransformer(model_name)
self._embedding_dimension = self.model.get_sentence_embedding_dimension()
logger.info(f"Initialized SentenceTransformer model '{model_name}' with dimension {self._embedding_dimension}")
except Exception as e:
logger.error(f"Failed to load SentenceTransformer model {model_name}: {e}")
raise
async def embed_documents(self, texts: List[str]) -> List[List[float]]:
# For CPU-bound tasks like local model inference, consider running in a thread pool
# to prevent blocking the event loop for long operations.
# For simplicity here, we'll call it directly.
return self.model.encode(texts).tolist()
async def embed_query(self, text: str) -> List[float]:
return self.model.encode(text).tolist()
class ThirdPartyAPIEmbeddingModel(BaseEmbeddingModel):
def __init__(self, model_name: str, api_endpoint: str, headers: Dict[str, str], payload_template: Dict[str, Any], embedding_dimension: int):
super().__init__(model_name)
self.api_endpoint = api_endpoint
self.headers = headers
self.payload_template = payload_template
self._embedding_dimension = embedding_dimension
self.session = None # aiohttp client session will be initialized on first use or in a context manager
logger.info(f"Initialized ThirdPartyAPIEmbeddingModel '{model_name}' for async calls to {api_endpoint} with dimension {embedding_dimension}")
async def _get_session(self):
"""Lazily create or return the aiohttp client session."""
if self.session is None or self.session.closed:
self.session = aiohttp.ClientSession()
return self.session
async def close_session(self):
"""Explicitly close the aiohttp client session."""
if self.session and not self.session.closed:
await self.session.close()
self.session = None
logger.info(f"Closed aiohttp session for model {self.model_name}")
async def embed_documents(self, texts: List[str]) -> List[List[float]]:
"""Asynchronously embeds a list of texts using the third-party API."""
session = await self._get_session()
embeddings = []
tasks = []
for text in texts:
payload = self.payload_template.copy()
if 'input' in payload:
payload['input'] = text
elif 'texts' in payload:
payload['texts'] = [text]
else:
raise ValueError("Payload template does not contain expected text input key.")
tasks.append(self._make_api_request(session, payload))
results = await asyncio.gather(*tasks, return_exceptions=True)
for i, res in enumerate(results):
if isinstance(res, Exception):
logger.error(f"Error embedding text '{texts[i][:50]}...': {res}")
# Depending on your error handling strategy, you might:
# - Append None or an empty list
# - Re-raise the exception to stop processing
# - Log and skip, then continue
embeddings.append([0.0] * self.embedding_dimension) # Append dummy embedding or handle failure
else:
embeddings.append(res)
return embeddings
async def _make_api_request(self, session: aiohttp.ClientSession, payload: Dict[str, Any]) -> List[float]:
"""Helper to make an asynchronous API request and extract embedding."""
try:
async with session.post(self.api_endpoint, headers=self.headers, json=payload) as response:
response.raise_for_status() # Raise an exception for HTTP errors (4xx, 5xx)
api_response = await response.json()
# Adjust this based on your API's actual response structure
if "data" in api_response and len(api_response["data"]) > 0 and "embedding" in api_response["data"][0]:
embedding = api_response["data"][0]["embedding"]
if len(embedding) != self.embedding_dimension:
logger.warning(f"API returned embedding of dimension {len(embedding)}, but expected {self.embedding_dimension} for model {self.model_name}. Adjusting config might be needed.")
return embedding
elif "embeddings" in api_response and isinstance(api_response["embeddings"], list) and api_response["embeddings"]:
embedding = api_response["embeddings"][0]
if len(embedding) != self.embedding_dimension:
logger.warning(f"API returned embedding of dimension {len(embedding)}, but expected {self.embedding_dimension} for model {self.model_name}. Adjusting config might be needed.")
return embedding
else:
raise ValueError(f"Unexpected API response structure: {api_response}")
except aiohttp.ClientError as e:
raise ConnectionError(f"API request failed: {e}") from e
except ValueError as e:
raise ValueError(f"Error processing API response: {e}") from e
async def embed_query(self, text: str) -> List[float]:
"""Asynchronously embeds a single query text."""
results = await self.embed_documents([text])
if results:
return results[0]
return [] # Or raise an error if embedding a query must always succeed
# --- Embedding Model Configuration ---
EMBEDDING_MODEL_CONFIGS: Dict[str, Dict[str, Any]] = {
"MiniLM": { # Example for a local Sentence Transformer model
"type": "sentence_transformer",
"model_name": "sentence-transformers/all-MiniLM-L6-v2"
},
"bge-m3": { # Example for a third-party API model
"type": "third_party_api",
"model_name": "bge-m3",
"api_endpoint": "https://api.qhaigc.net/v1/embeddings",
"headers": {
"Content-Type": "application/json",
"Authorization": f"Bearer {os.getenv('rag_api_key')}"
},
"payload_template": {
"model": "bge-m3",
"input": ""
},
"embedding_dimension": 1024
},
"OpenAI-Ada-002": {
"type": "third_party_api",
"model_name": "text-embedding-ada-002",
"api_endpoint": "https://api.openai.com/v1/embeddings",
"headers": {
"Content-Type": "application/json",
"Authorization": f"Bearer {os.getenv('OPENAI_API_KEY')}" # Ensure OPENAI_API_KEY is set
},
"payload_template": {
"model": "text-embedding-ada-002",
"input": "" # Text will be injected here
},
"embedding_dimension": 1536
},
"OpenAI-Embedding-3-Small": {
"type": "third_party_api",
"model_name": "text-embedding-3-small",
"api_endpoint": "https://api.openai.com/v1/embeddings",
"headers": {
"Content-Type": "application/json",
"Authorization": f"Bearer {os.getenv('OPENAI_API_KEY')}"
},
"payload_template": {
"model": "text-embedding-3-small",
"input": "",
# "dimensions": 512 # Optional: uncomment if you want a specific output dimension
},
"embedding_dimension": 1536 # Default max dimension for text-embedding-3-small
},
}