282 lines
10 KiB
Python
282 lines
10 KiB
Python
import json
|
|
from typing import Union, AsyncGenerator, List, Optional, Dict, Any
|
|
import logging
|
|
|
|
import httpx
|
|
|
|
from .datatypes import LLMBackend, LLMMessage
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
class LLMClient:
|
|
"""Client for interacting with LLM APIs"""
|
|
backend: LLMBackend
|
|
embedding_backend: LLMBackend
|
|
timeout: float
|
|
|
|
def __init__(self, backend: LLMBackend, embedding_backend: Optional[LLMBackend], timeout: float = 30.0):
|
|
"""Initialize the LLM client
|
|
|
|
Args:
|
|
backend: LLM backend configuration containing base_url, api_token, and model
|
|
"""
|
|
self.backend = backend
|
|
self.embedding_backend = embedding_backend if embedding_backend else backend
|
|
self.timeout = timeout
|
|
|
|
async def get_embedding(self, text: str) -> List[float]:
|
|
"""Get embedding for text
|
|
|
|
Args:
|
|
text: Text to get embedding for
|
|
|
|
Returns:
|
|
List of float values representing the embedding vector
|
|
"""
|
|
try:
|
|
# Use provided model or fall back to backend model
|
|
|
|
request_params = {
|
|
"model": self.embedding_backend["model"],
|
|
"prompt": text
|
|
}
|
|
|
|
headers = {
|
|
"Content-Type": "application/json"
|
|
}
|
|
if len(self.embedding_backend["api_token"]):
|
|
headers['Authorization'] = f"Bearer {self.embedding_backend['api_token']}"
|
|
|
|
async with httpx.AsyncClient(timeout=self.timeout) as client:
|
|
url = f"{self.embedding_backend['base_url']}/embeddings"
|
|
|
|
response = await client.post(
|
|
url,
|
|
headers=headers,
|
|
json=request_params,
|
|
)
|
|
response.raise_for_status()
|
|
|
|
response_data = response.json()
|
|
|
|
# Extract embedding from response
|
|
if "embedding" in response_data and response_data["embedding"]:
|
|
return response_data["embedding"]
|
|
else:
|
|
logger.error("No embedding data in response")
|
|
return []
|
|
|
|
except httpx.HTTPStatusError as e:
|
|
logger.error(f"HTTP error getting embedding: {e.response.status_code} - {e.response.text}")
|
|
return []
|
|
|
|
except httpx.RequestError as e:
|
|
logger.error(f"Request error getting embedding: {str(e)}")
|
|
return []
|
|
|
|
except Exception as e:
|
|
logger.error(f"Error getting embedding: {str(e)}")
|
|
return []
|
|
|
|
async def get_response(self, messages: List[LLMMessage], stream: Optional[bool]) -> AsyncGenerator[dict[str, Any] | str, Any]:
|
|
"""Get response from the LLM
|
|
|
|
Args:
|
|
messages: List of messages to send to the LLM
|
|
stream: Whether to stream responses by default
|
|
|
|
Returns:
|
|
Either a string response or an async generator for streaming
|
|
"""
|
|
try:
|
|
|
|
stream = stream if stream else False
|
|
|
|
# Prepare the request parameters
|
|
request_params = {
|
|
"model": self.backend["model"],
|
|
"messages": messages,
|
|
"stream": stream,
|
|
}
|
|
# Prepare headers
|
|
headers = {
|
|
"Content-Type": "application/json"
|
|
}
|
|
if len(self.backend["api_token"]):
|
|
headers['Authorization'] = f"Bearer {self.backend['api_token']}"
|
|
|
|
logger.info(headers)
|
|
logger.info(request_params)
|
|
|
|
# Create httpx client
|
|
async with httpx.AsyncClient(timeout=self.timeout) as client:
|
|
url = f"{self.backend['base_url']}/chat"
|
|
|
|
if stream:
|
|
# Stream the response
|
|
async with client.stream(
|
|
"POST",
|
|
url,
|
|
headers=headers,
|
|
json=request_params,
|
|
) as response:
|
|
response.raise_for_status()
|
|
|
|
async for line in response.aiter_lines():
|
|
line = line.strip()
|
|
|
|
# Skip empty lines and non-data lines
|
|
if not line or not line.startswith("data: "):
|
|
continue
|
|
|
|
# Remove "data: " prefix
|
|
data = line[6:]
|
|
|
|
# Check for stream end
|
|
if data == "[DONE]":
|
|
break
|
|
|
|
try:
|
|
# Parse JSON chunk
|
|
chunk_data = json.loads(data)
|
|
|
|
if "choices" in chunk_data and chunk_data["choices"]:
|
|
choice = chunk_data["choices"][0]
|
|
delta = choice.get("delta", {})
|
|
|
|
# Handle reasoning content (for models that support it)
|
|
if "reasoning_content" in delta and delta["reasoning_content"]:
|
|
yield {'reasoning': delta["reasoning_content"]} # type: ignore
|
|
|
|
# Handle regular content
|
|
if "content" in delta and delta["content"]:
|
|
yield {'content': delta["content"]} # type: ignore
|
|
|
|
except json.JSONDecodeError:
|
|
# Skip malformed JSON chunks
|
|
continue
|
|
else:
|
|
# Non-streaming response
|
|
response = await client.post(
|
|
url,
|
|
headers=headers,
|
|
json=request_params,
|
|
)
|
|
response.raise_for_status()
|
|
|
|
response_data = response.json()
|
|
content = ""
|
|
|
|
# if "message" in response_data and response_data["message"]:
|
|
# content = response_data["message"][0]['content']
|
|
content = response_data["message"]['content']
|
|
logger.info(response_data)
|
|
# FIX: Yield as dictionary to match streaming format
|
|
if content:
|
|
logger.info(content)
|
|
yield {'content': content} # type: ignore
|
|
|
|
except httpx.HTTPStatusError as e:
|
|
error_msg = f"HTTP error getting LLM response: {e.response.status_code} - {e.response.text}"
|
|
logger.error(error_msg)
|
|
yield ""
|
|
|
|
except httpx.RequestError as e:
|
|
error_msg = f"Request error getting LLM response: {str(e)}"
|
|
logger.error(error_msg)
|
|
yield ""
|
|
|
|
except Exception as e:
|
|
error_msg = f"Error getting LLM response: {str(e)}"
|
|
logger.error(error_msg)
|
|
yield ""
|
|
|
|
async def get_structured_response(self, messages: List[LLMMessage], json_format: Dict[str, Any]) -> Dict[str, Any]:
|
|
"""Get structured JSON response from the LLM using a JSON schema
|
|
|
|
Args:
|
|
messages: List of messages to send to the LLM
|
|
json_format: JSON schema for structured output
|
|
|
|
Returns:
|
|
Parsed JSON response as dictionary
|
|
|
|
Raises:
|
|
ValueError: If the response is not valid JSON
|
|
HTTPError: If the API request fails
|
|
"""
|
|
try:
|
|
# Prepare the request parameters with format
|
|
request_params = {
|
|
"model": self.backend["model"],
|
|
"messages": messages,
|
|
"format": json_format, # Ollama's structured output parameter
|
|
"stream": False,
|
|
}
|
|
|
|
# Prepare headers
|
|
headers = {
|
|
"Content-Type": "application/json"
|
|
}
|
|
if len(self.backend["api_token"]):
|
|
headers['Authorization'] = f"Bearer {self.backend['api_token']}"
|
|
|
|
logger.info("Structured request headers: %s", headers)
|
|
logger.info("Structured request params: %s", request_params)
|
|
|
|
# Create httpx client
|
|
async with httpx.AsyncClient(timeout=self.timeout) as client:
|
|
url = f"{self.backend['base_url']}/chat"
|
|
|
|
# Non-streaming response only
|
|
response = await client.post(
|
|
url,
|
|
headers=headers,
|
|
json=request_params,
|
|
)
|
|
response.raise_for_status()
|
|
|
|
response_data = response.json()
|
|
logger.info("Structured response data: %s", response_data)
|
|
|
|
# Extract content from response
|
|
if "message" not in response_data or not response_data["message"]:
|
|
raise ValueError("No message in response")
|
|
|
|
content = response_data["message"].get('content', '')
|
|
|
|
if not content:
|
|
raise ValueError("Empty content in structured response")
|
|
|
|
# Parse JSON content
|
|
try:
|
|
structured_data = json.loads(content)
|
|
logger.info("Parsed structured data: %s", structured_data)
|
|
return structured_data
|
|
|
|
except json.JSONDecodeError as e:
|
|
logger.error("Failed to parse structured response as JSON: %s", content)
|
|
raise ValueError(f"Response is not valid JSON: {e}")
|
|
|
|
except httpx.HTTPStatusError as e:
|
|
error_msg = f"HTTP error getting structured LLM response: {e.response.status_code} - {e.response.text}"
|
|
logger.error(error_msg)
|
|
raise
|
|
|
|
except httpx.RequestError as e:
|
|
error_msg = f"Request error getting structured LLM response: {str(e)}"
|
|
logger.error(error_msg)
|
|
raise
|
|
|
|
except Exception as e:
|
|
error_msg = f"Error getting structured LLM response: {str(e)}"
|
|
logger.error(error_msg)
|
|
raise
|
|
|
|
|
|
async def _empty_async_generator() -> AsyncGenerator[str, None]:
|
|
"""Helper function for empty async generator"""
|
|
if False:
|
|
yield ""
|