Spaces:
Sleeping
Sleeping
File size: 11,223 Bytes
5e1a30c |
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 |
"""
Base adapter implementation for LLM providers.
This module provides common functionality shared by all LLM adapters,
including error handling, retry logic, and format conversion utilities.
Architecture Notes:
- All LLM-specific logic must be in adapters, not in the orchestrator
- Adapters handle authentication, format conversion, and error mapping
- Each provider's peculiarities are hidden behind the unified interface
"""
import time
import logging
from typing import Dict, Any, Optional, List, Iterator
from abc import abstractmethod
import json
from ..base import LLMAdapter, LLMError, GenerationParams, ConfigurableComponent
logger = logging.getLogger(__name__)
class RateLimitError(LLMError):
"""Rate limit exceeded error."""
pass
class AuthenticationError(LLMError):
"""Authentication failed error."""
pass
class ModelNotFoundError(LLMError):
"""Model not found error."""
pass
class BaseLLMAdapter(LLMAdapter, ConfigurableComponent):
"""
Base implementation for LLM adapters with common functionality.
Provides:
- Retry logic with exponential backoff
- Common error handling and mapping
- Response validation
- Metrics collection
- Configuration management
Subclasses must implement:
- _make_request: Provider-specific API call
- _parse_response: Convert provider response to text
- _get_provider_name: Return provider name
- _validate_model: Check if model exists
"""
def __init__(self,
model_name: str,
config: Optional[Dict[str, Any]] = None,
max_retries: int = 3,
retry_delay: float = 1.0):
"""
Initialize base LLM adapter.
Args:
model_name: Name of the model to use
config: Provider-specific configuration
max_retries: Maximum number of retry attempts
retry_delay: Initial delay between retries (seconds)
"""
super().__init__(config)
self.model_name = model_name
self.max_retries = max_retries
self.retry_delay = retry_delay
self._request_count = 0
self._total_tokens = 0
self._last_request_time = 0
def generate(self, prompt: str, params: GenerationParams) -> str:
"""
Generate a response with retry logic and error handling.
Args:
prompt: The prompt to send to the LLM
params: Generation parameters
Returns:
Generated text response
Raises:
LLMError: If generation fails after retries
"""
if not prompt.strip():
raise ValueError("Prompt cannot be empty")
# Track request timing
start_time = time.time()
self._last_request_time = start_time
# Attempt generation with retries
last_error = None
for attempt in range(self.max_retries):
try:
# Make provider-specific request
response = self._make_request(prompt, params)
# Parse provider response
text = self._parse_response(response)
# Validate response
if not text or not text.strip():
raise LLMError("Empty response from LLM")
# Update metrics
self._request_count += 1
if 'usage' in response:
self._total_tokens += response['usage'].get('total_tokens', 0)
# Log success
elapsed = time.time() - start_time
logger.info(f"{self._get_provider_name()} generation completed in {elapsed:.2f}s")
return text
except RateLimitError as e:
# Rate limit - use exponential backoff
last_error = e
if attempt < self.max_retries - 1:
delay = self.retry_delay * (2 ** attempt)
logger.warning(f"Rate limit hit, retrying in {delay}s...")
time.sleep(delay)
else:
logger.error(f"Rate limit exceeded after {self.max_retries} attempts")
except AuthenticationError as e:
# Authentication errors are not retryable
logger.error(f"Authentication failed: {str(e)}")
raise
except Exception as e:
# Other errors - retry with backoff
last_error = e
if attempt < self.max_retries - 1:
delay = self.retry_delay * (2 ** attempt)
logger.warning(f"Generation failed: {str(e)}, retrying in {delay}s...")
time.sleep(delay)
else:
logger.error(f"Generation failed after {self.max_retries} attempts: {str(e)}")
# All retries exhausted
raise LLMError(f"Generation failed after {self.max_retries} attempts: {str(last_error)}")
def generate_streaming(self, prompt: str, params: GenerationParams) -> Iterator[str]:
"""
Generate a streaming response.
Default implementation calls generate() and yields the result.
Subclasses should override for true streaming support.
Args:
prompt: The prompt to send to the LLM
params: Generation parameters
Yields:
Generated text chunks
"""
# Default: yield full response as single chunk
# Subclasses should override for true streaming
logger.warning(f"{self._get_provider_name()} adapter using simulated streaming")
yield self.generate(prompt, params)
def get_model_info(self) -> Dict[str, Any]:
"""
Get information about the model and provider.
Returns:
Dictionary with model information
"""
return {
'provider': self._get_provider_name(),
'model': self.model_name,
'supports_streaming': self._supports_streaming(),
'max_tokens': self._get_max_tokens(),
'requests_made': self._request_count,
'total_tokens_used': self._total_tokens,
'configuration': self._get_safe_config()
}
def validate_connection(self) -> bool:
"""
Validate the connection to the LLM provider.
Returns:
True if connection is valid
Raises:
LLMError: If connection validation fails
"""
try:
# Validate model exists
if not self._validate_model():
raise ModelNotFoundError(f"Model '{self.model_name}' not found")
# Make a minimal test request
test_prompt = "Hello"
test_params = GenerationParams(max_tokens=10, temperature=0)
response = self.generate(test_prompt, test_params)
return bool(response)
except Exception as e:
logger.error(f"Connection validation failed: {str(e)}")
raise LLMError(f"Failed to validate {self._get_provider_name()} connection: {str(e)}")
# Abstract methods that subclasses must implement
@abstractmethod
def _make_request(self, prompt: str, params: GenerationParams) -> Dict[str, Any]:
"""
Make a request to the LLM provider.
Args:
prompt: The prompt to send
params: Generation parameters
Returns:
Raw response from provider
Raises:
Provider-specific exceptions
"""
pass
@abstractmethod
def _parse_response(self, response: Dict[str, Any]) -> str:
"""
Parse the provider response to extract generated text.
Args:
response: Raw response from provider
Returns:
Generated text
"""
pass
@abstractmethod
def _get_provider_name(self) -> str:
"""Return the provider name (e.g., 'Ollama', 'OpenAI')."""
pass
@abstractmethod
def _validate_model(self) -> bool:
"""Check if the configured model exists/is available."""
pass
# Optional methods that subclasses can override
def _supports_streaming(self) -> bool:
"""Whether this adapter supports true streaming."""
return False
def _get_max_tokens(self) -> Optional[int]:
"""Get maximum token limit for this model."""
return None
def _get_safe_config(self) -> Dict[str, Any]:
"""Get configuration with sensitive data removed."""
safe_config = self.config.copy()
# Remove sensitive keys
sensitive_keys = ['api_key', 'token', 'secret', 'password']
for key in sensitive_keys:
if key in safe_config:
safe_config[key] = '***'
return safe_config
# Utility methods for subclasses
def _handle_provider_error(self, error: Exception) -> None:
"""
Map provider-specific errors to standard errors.
Subclasses should override to handle specific error types.
Args:
error: Provider-specific error
Raises:
Appropriate LLMError subclass
"""
error_msg = str(error).lower()
if 'rate limit' in error_msg or '429' in error_msg:
raise RateLimitError(str(error))
elif 'unauthorized' in error_msg or '401' in error_msg or 'api key' in error_msg:
raise AuthenticationError(str(error))
elif 'not found' in error_msg or '404' in error_msg:
raise ModelNotFoundError(str(error))
else:
raise LLMError(f"Provider error: {str(error)}")
def _prepare_messages(self, prompt: str) -> List[Dict[str, str]]:
"""
Prepare messages in chat format.
Many providers expect chat-style message format.
Args:
prompt: The prompt text
Returns:
List of message dictionaries
"""
return [{"role": "user", "content": prompt}]
def _extract_content(self, message: Dict[str, Any]) -> str:
"""
Extract content from a message object.
Handles various message formats from different providers.
Args:
message: Message object
Returns:
Extracted content string
"""
# Try common content locations
if isinstance(message, str):
return message
elif isinstance(message, dict):
return (message.get('content') or
message.get('text') or
message.get('message') or
str(message))
else:
return str(message) |