UDHOV's picture
Sync from GitHub via hub-sync
41de262 verified
Raw
History Blame Contribute Delete
15.7 kB
"""
LLM Client
Centralized LLM API calls with support for both Ollama (local) and external APIs.
Uses retry logic and circuit breaker pattern.
"""
import logging
import os
import time
from dataclasses import dataclass
from typing import Any, cast
try:
from dotenv import load_dotenv
load_dotenv()
except ImportError:
pass
import requests
from src.reasoning.utils.api_llm_client import InvalidAPIKeyError
from src.reasoning.utils.config_loader import ConfigLoader
from src.reasoning.utils.json_parser import safe_json_parse
logger = logging.getLogger(__name__)
@dataclass
class LLMResponse:
"""Structured LLM response."""
text: str
raw_response: dict[str, Any]
latency_ms: float
success: bool
error: str | None = None
class CircuitBreaker:
"""Circuit breaker pattern to prevent repeated failing calls."""
def __init__(self, failure_threshold: int = 5, cooldown_seconds: int = 60) -> None:
self.failure_threshold = failure_threshold
self.cooldown_seconds = cooldown_seconds
self.failures = 0
self.last_failure_time: float | None = None
self.is_open = False
def record_success(self) -> None:
"""Record a successful call."""
self.failures = 0
self.is_open = False
def record_failure(self) -> None:
"""Record a failed call."""
self.failures += 1
self.last_failure_time = time.time()
if self.failures >= self.failure_threshold:
self.is_open = True
logger.warning(f"Circuit breaker opened after {self.failures} consecutive failures")
def can_proceed(self) -> bool:
"""Check if calls can proceed."""
if not self.is_open:
return True
if self.last_failure_time is None:
return True
if time.time() - self.last_failure_time > self.cooldown_seconds:
logger.info("Circuit breaker cooling down, allowing attempt")
self.is_open = False
return True
return False
class LLMClient:
"""
Unified LLM client with Dynamic Provider Detection.
Automatically selects the best available provider based on environment keys.
"""
def __init__(
self,
config_path: str = "config/settings.yaml",
max_retries: int = 3,
timeout: int = 120,
) -> None:
self.config = ConfigLoader(config_path)
self.max_retries = max_retries
self.timeout = timeout
self.circuit_breaker = CircuitBreaker()
self._api_client: Any = None
self.active_profile: str = "unknown"
self._ollama_cache: tuple[bool, float] | None = None
self._OLLAMA_CACHE_TTL = 30
self._select_provider()
def _select_provider(self) -> None:
"""Selects and initializes the LLM provider based on config and env."""
requested_provider = str(self.config.get("llm.provider", "auto"))
profiles = cast(dict[str, Any], self.config.get("llm.profiles", {}))
# Always initialize Ollama as fallback (even if not selected)
ollama_profile = profiles.get("ollama")
if ollama_profile:
url = ollama_profile.get("url", "http://localhost:11434/api/generate")
try:
tags_url = url.replace("/api/generate", "/api/tags")
resp = requests.get(tags_url, timeout=2)
if resp.status_code == 200:
self._ollama_url = url
self._ollama_model = ollama_profile.get("model", "llama3")
logger.info(f"LLM Client: Ollama available as fallback at {url}")
except Exception:
logger.debug(f"Ollama not available at {url}")
if requested_provider == "auto":
# Priority-based auto-detection (skip disabled/working profiles)
priority_list = [
"openrouter",
"openai",
"groq",
] # HuggingFace disabled for now
for profile_name in priority_list:
profile = profiles.get(profile_name)
# Skip if disabled in comments or missing
if profile and profile.get("type") != "disabled" and self._try_init_profile(profile_name, profile):
return
# If API providers fail, use Ollama fallback
if hasattr(self, "_ollama_url") and self._ollama_url:
self.provider = "ollama"
self.active_profile = "ollama"
logger.info("LLM Client: Using Ollama fallback")
return
# If all else fails, use a fallback error state
logger.critical("No valid LLM providers detected. Set an API key in .env")
else:
# Explicit selection
profile = profiles.get(requested_provider)
if not self._try_init_profile(requested_provider, profile):
logger.error(f"Failed to initialize requested provider: {requested_provider}")
def _try_init_profile(self, name: str, profile: dict | None) -> bool:
"""Attempts to initialize a specific provider profile."""
if not profile:
return False
profile_type = profile.get("type")
if profile_type == "api":
env_key = profile.get("env_key", "")
api_key = os.getenv(env_key, "")
if api_key:
try:
from src.reasoning.utils.api_llm_client import (
APIConfig,
APILLMClient,
)
config = APIConfig(
endpoint=profile.get("endpoint", ""),
api_key=api_key,
model=profile.get("model", ""),
timeout=profile.get("timeout", self.timeout),
)
self._api_client = APILLMClient(config)
self.provider = "api"
self.active_profile = name
self.timeout = config.timeout
logger.info(f"LLM Client: Auto-detected '{name}' via {env_key}")
return True
except Exception as e:
logger.warning(f"Failed to init {name} profile: {e}")
return False
elif profile_type == "ollama":
url = profile.get("url", "http://localhost:11434/api/generate")
# Minimal reachability check
try:
# We don't want to wait too long for a check
resp = requests.get(url.replace("/api/generate", "/api/tags"), timeout=2)
if resp.status_code == 200:
self._ollama_url = url
self._ollama_model = profile.get("model", "llama3")
self.provider = "ollama"
self.active_profile = name
self.timeout = profile.get("timeout", self.timeout)
logger.info(f"LLM Client: Auto-detected local Ollama at {url}")
return True
except Exception:
logger.debug(f"Ollama profile '{name}' not reachable at {url}")
return False
return False
def generate(
self,
prompt: str,
format_json: bool = False,
temperature: float = 0.0,
custom_model: str | None = None,
llm_api_key: str | None = None,
) -> LLMResponse:
"""Generates a response with automatic fallback between providers."""
# User-provided key takes priority — send to OpenRouter endpoint
if llm_api_key:
try:
profiles = cast(dict[str, Any], self.config.get("llm.profiles", {}))
openrouter_profile = profiles.get("openrouter", {})
if not openrouter_profile:
raise ValueError("OpenRouter profile not found in config")
from src.reasoning.utils.api_llm_client import (
APIConfig,
APILLMClient,
)
or_config = APIConfig(
endpoint=openrouter_profile.get("endpoint", ""),
api_key=llm_api_key,
model=openrouter_profile.get("model", ""),
timeout=openrouter_profile.get("timeout", self.timeout),
)
api_client = APILLMClient(or_config)
result = api_client.generate(prompt, format_json, temperature, custom_model)
if result["success"]:
return LLMResponse(
text=result["text"],
raw_response=result["raw_response"],
latency_ms=result["latency_ms"],
success=True,
error=None,
)
return LLMResponse(
text="",
raw_response={},
latency_ms=result.get("latency_ms", 0.0),
success=False,
error=result.get("error", "Generation failed"),
)
except InvalidAPIKeyError as e:
logger.error("User-provided OpenRouter key failed: %s", str(e))
return LLMResponse(
text="",
raw_response={},
latency_ms=0.0,
success=False,
error="invalid_api_key: The OpenRouter API key you provided is invalid or expired. "
"Please check your key in Settings and try again.",
)
except Exception as e:
logger.error("User-provided OpenRouter key failed: %s", str(e))
return LLMResponse(
text="",
raw_response={},
latency_ms=0.0,
success=False,
error=str(e),
)
# Verify API key is still valid before trying API provider
profiles = cast(dict[str, Any], self.config.get("llm.profiles", {}))
active_profile = profiles.get(self.active_profile)
api_key_valid = False
if active_profile and active_profile.get("type") == "api":
env_key = active_profile.get("env_key", "")
api_key_valid = bool(os.getenv(env_key, ""))
# Try API provider first (if set and key still valid)
if self.provider == "api" and api_key_valid and self.circuit_breaker.can_proceed():
try:
result = self._api_client.generate(prompt, format_json, temperature, custom_model)
if result["success"]:
self.circuit_breaker.record_success()
return LLMResponse(
text=result["text"],
raw_response=result["raw_response"],
latency_ms=result["latency_ms"],
success=True,
error=None,
)
# API returned success=False (e.g. max retries exceeded inside _api_client)
self.circuit_breaker.record_failure()
logger.warning(f"API failed: {result.get('error')}")
except InvalidAPIKeyError as e:
self.circuit_breaker.record_failure()
logger.error("System API key invalid: %s", str(e))
except Exception as e:
# Catch connection errors, timeouts, or 429/500 errors from httpx
self.circuit_breaker.record_failure()
logger.error(f"API client exception: {str(e)}")
# Try other API profiles before falling back to Ollama
profiles = cast(dict[str, Any], self.config.get("llm.profiles", {}))
for profile_name in ["openai", "openrouter"]: # HuggingFace disabled
profile = profiles.get(profile_name)
if profile and profile_name != self.active_profile and self._try_init_profile(profile_name, profile):
logger.info(f"Trying fallback profile: {profile_name}")
try:
result = self._api_client.generate(prompt, format_json, temperature, custom_model)
if result["success"]:
self.circuit_breaker.record_success()
return LLMResponse(
text=result["text"],
raw_response=result["raw_response"],
latency_ms=result["latency_ms"],
success=True,
error=None,
)
logger.warning(f"Fallback {profile_name} also failed")
except Exception as e:
logger.warning(f"Fallback {profile_name} exception: {str(e)}")
# Fallback to Ollama
if self._ollama_available():
logger.info("Using Ollama fallback")
return self._generate_ollama(prompt, format_json, temperature, custom_model)
# No providers available
return LLMResponse("", {}, 0.0, False, "All providers failed")
def _ollama_available(self) -> bool:
"""Check if Ollama is available, with 30-second caching."""
now = time.time()
if self._ollama_cache:
available, ts = self._ollama_cache
if now - ts < self._OLLAMA_CACHE_TTL:
return available
# Cache miss or expired — re-check
available = hasattr(self, "_ollama_url") and bool(self._ollama_url)
self._ollama_cache = (available, now)
return available
def _generate_ollama(
self,
prompt: str,
format_json: bool,
temperature: float,
custom_model: str | None,
) -> LLMResponse:
"""Generate using Ollama."""
model = custom_model or self._ollama_model
payload = {
"model": model,
"prompt": prompt,
"stream": False,
"options": {"temperature": temperature},
}
if format_json:
payload["format"] = "json"
last_error = None
for attempt in range(self.max_retries):
start_time = time.perf_counter()
# Use longer timeout for Ollama (30 minutes for CPU inference)
ollama_timeout = 1800
try:
response = requests.post(self._ollama_url, json=payload, timeout=ollama_timeout)
response.raise_for_status()
result = response.json()
latency_ms = (time.perf_counter() - start_time) * 1000
self.circuit_breaker.record_success()
return LLMResponse(result.get("response", ""), result, latency_ms, True)
except Exception as e:
last_error = str(e)
if attempt < self.max_retries - 1:
time.sleep((2**attempt) * 0.5)
self.circuit_breaker.record_failure()
return LLMResponse("", {}, 0.0, False, last_error or "Max retries exceeded")
def generate_json(
self,
prompt: str,
temperature: float = 0.0,
default: dict[str, Any] | None = None,
llm_api_key: str | None = None,
) -> dict[str, Any]:
"""Generate and parse JSON response."""
response = self.generate(prompt, format_json=True, temperature=temperature, llm_api_key=llm_api_key)
if not response.success:
return default or {}
return safe_json_parse(response.text, default)