d49e8201b4
* feat: add structured output to /reflect * feat: add structured output to /reflect * imrpove * add max_toksn * fix rust client * fix rust client * fix rust client * try fix * try fix * no stricts
854 lines
38 KiB
Python
854 lines
38 KiB
Python
"""
|
|
LLM wrapper for unified configuration across providers.
|
|
"""
|
|
|
|
import asyncio
|
|
import json
|
|
import logging
|
|
import os
|
|
import time
|
|
from typing import Any
|
|
|
|
import httpx
|
|
from google import genai
|
|
from google.genai import errors as genai_errors
|
|
from google.genai import types as genai_types
|
|
from openai import APIConnectionError, APIStatusError, AsyncOpenAI, LengthFinishReasonError
|
|
|
|
from ..config import (
|
|
DEFAULT_LLM_MAX_CONCURRENT,
|
|
DEFAULT_LLM_TIMEOUT,
|
|
ENV_LLM_MAX_CONCURRENT,
|
|
ENV_LLM_TIMEOUT,
|
|
)
|
|
|
|
# Seed applied to every Groq request for deterministic behavior.
|
|
DEFAULT_LLM_SEED = 4242
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
# Disable httpx logging
|
|
logging.getLogger("httpx").setLevel(logging.WARNING)
|
|
|
|
# Global semaphore to limit concurrent LLM requests across all instances
|
|
# Set HINDSIGHT_API_LLM_MAX_CONCURRENT=1 for local LLMs (LM Studio, Ollama)
|
|
_llm_max_concurrent = int(os.getenv(ENV_LLM_MAX_CONCURRENT, str(DEFAULT_LLM_MAX_CONCURRENT)))
|
|
_global_llm_semaphore = asyncio.Semaphore(_llm_max_concurrent)
|
|
|
|
|
|
class OutputTooLongError(Exception):
|
|
"""
|
|
Bridge exception raised when LLM output exceeds token limits.
|
|
|
|
This wraps provider-specific errors (e.g., OpenAI's LengthFinishReasonError)
|
|
to allow callers to handle output length issues without depending on
|
|
provider-specific implementations.
|
|
"""
|
|
|
|
pass
|
|
|
|
|
|
class LLMProvider:
|
|
"""
|
|
Unified LLM provider.
|
|
|
|
Supports OpenAI, Groq, Ollama (OpenAI-compatible), and Gemini.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
provider: str,
|
|
api_key: str,
|
|
base_url: str,
|
|
model: str,
|
|
reasoning_effort: str = "low",
|
|
):
|
|
"""
|
|
Initialize LLM provider.
|
|
|
|
Args:
|
|
provider: Provider name ("openai", "groq", "ollama", "gemini", "anthropic", "lmstudio").
|
|
api_key: API key.
|
|
base_url: Base URL for the API.
|
|
model: Model name.
|
|
reasoning_effort: Reasoning effort level for supported providers.
|
|
"""
|
|
self.provider = provider.lower()
|
|
self.api_key = api_key
|
|
self.base_url = base_url
|
|
self.model = model
|
|
self.reasoning_effort = reasoning_effort
|
|
|
|
# Validate provider
|
|
valid_providers = ["openai", "groq", "ollama", "gemini", "anthropic", "lmstudio"]
|
|
if self.provider not in valid_providers:
|
|
raise ValueError(f"Invalid LLM provider: {self.provider}. Must be one of: {', '.join(valid_providers)}")
|
|
|
|
# Set default base URLs
|
|
if not self.base_url:
|
|
if self.provider == "groq":
|
|
self.base_url = "https://api.groq.com/openai/v1"
|
|
elif self.provider == "ollama":
|
|
self.base_url = "http://localhost:11434/v1"
|
|
elif self.provider == "lmstudio":
|
|
self.base_url = "http://localhost:1234/v1"
|
|
|
|
# Validate API key (not needed for ollama or lmstudio)
|
|
if self.provider not in ("ollama", "lmstudio") and not self.api_key:
|
|
raise ValueError(f"API key not found for {self.provider}")
|
|
|
|
# Get timeout config (set HINDSIGHT_API_LLM_TIMEOUT for local LLMs that need longer timeouts)
|
|
self.timeout = float(os.getenv(ENV_LLM_TIMEOUT, str(DEFAULT_LLM_TIMEOUT)))
|
|
|
|
# Create client based on provider
|
|
self._client = None
|
|
self._gemini_client = None
|
|
self._anthropic_client = None
|
|
|
|
if self.provider == "gemini":
|
|
self._gemini_client = genai.Client(api_key=self.api_key)
|
|
elif self.provider == "anthropic":
|
|
from anthropic import AsyncAnthropic
|
|
|
|
# Only pass base_url if it's set (Anthropic uses default URL otherwise)
|
|
anthropic_kwargs = {"api_key": self.api_key}
|
|
if self.base_url:
|
|
anthropic_kwargs["base_url"] = self.base_url
|
|
if self.timeout:
|
|
anthropic_kwargs["timeout"] = self.timeout
|
|
self._anthropic_client = AsyncAnthropic(**anthropic_kwargs)
|
|
elif self.provider in ("ollama", "lmstudio"):
|
|
# Use dummy key if not provided for local
|
|
api_key = self.api_key or "local"
|
|
client_kwargs = {
|
|
"api_key": api_key,
|
|
"base_url": self.base_url,
|
|
"max_retries": 0
|
|
}
|
|
if self.timeout:
|
|
client_kwargs["timeout"] = self.timeout
|
|
self._client = AsyncOpenAI(**client_kwargs)
|
|
else:
|
|
# Only pass base_url if it's set (OpenAI uses default URL otherwise)
|
|
client_kwargs = {"api_key": self.api_key, "max_retries": 0}
|
|
if self.base_url:
|
|
client_kwargs["base_url"] = self.base_url
|
|
if self.timeout:
|
|
client_kwargs["timeout"] = self.timeout
|
|
self._client = AsyncOpenAI(**client_kwargs)
|
|
|
|
async def verify_connection(self) -> None:
|
|
"""
|
|
Verify that the LLM provider is configured correctly by making a simple test call.
|
|
|
|
Raises:
|
|
RuntimeError: If the connection test fails.
|
|
"""
|
|
try:
|
|
logger.info(
|
|
f"Verifying LLM: provider={self.provider}, model={self.model}, base_url={self.base_url or 'default'}..."
|
|
)
|
|
await self.call(
|
|
messages=[{"role": "user", "content": "Say 'ok'"}],
|
|
max_completion_tokens=100,
|
|
max_retries=2,
|
|
initial_backoff=0.5,
|
|
max_backoff=2.0,
|
|
)
|
|
# If we get here without exception, the connection is working
|
|
logger.info(f"LLM verified: {self.provider}/{self.model}")
|
|
except Exception as e:
|
|
raise RuntimeError(f"LLM connection verification failed for {self.provider}/{self.model}: {e}") from e
|
|
|
|
async def call(
|
|
self,
|
|
messages: list[dict[str, str]],
|
|
response_format: Any | None = None,
|
|
max_completion_tokens: int | None = None,
|
|
temperature: float | None = None,
|
|
scope: str = "memory",
|
|
max_retries: int = 10,
|
|
initial_backoff: float = 1.0,
|
|
max_backoff: float = 60.0,
|
|
skip_validation: bool = False,
|
|
strict_schema: bool = False,
|
|
) -> Any:
|
|
"""
|
|
Make an LLM API call with retry logic.
|
|
|
|
Args:
|
|
messages: List of message dicts with 'role' and 'content'.
|
|
response_format: Optional Pydantic model for structured output.
|
|
max_completion_tokens: Maximum tokens in response.
|
|
temperature: Sampling temperature (0.0-2.0).
|
|
scope: Scope identifier for tracking.
|
|
max_retries: Maximum retry attempts.
|
|
initial_backoff: Initial backoff time in seconds.
|
|
max_backoff: Maximum backoff time in seconds.
|
|
skip_validation: Return raw JSON without Pydantic validation.
|
|
strict_schema: Use strict JSON schema enforcement (OpenAI only). Guarantees all required fields.
|
|
|
|
Returns:
|
|
Parsed response if response_format is provided, otherwise text content.
|
|
|
|
Raises:
|
|
OutputTooLongError: If output exceeds token limits.
|
|
Exception: Re-raises API errors after retries exhausted.
|
|
"""
|
|
async with _global_llm_semaphore:
|
|
start_time = time.time()
|
|
|
|
# Handle Gemini provider separately
|
|
if self.provider == "gemini":
|
|
return await self._call_gemini(
|
|
messages, response_format, max_retries, initial_backoff, max_backoff, skip_validation, start_time
|
|
)
|
|
|
|
# Handle Anthropic provider separately
|
|
if self.provider == "anthropic":
|
|
return await self._call_anthropic(
|
|
messages, response_format, max_completion_tokens, max_retries, initial_backoff, max_backoff, skip_validation, start_time
|
|
)
|
|
|
|
# Handle Ollama with native API for structured output (better schema enforcement)
|
|
if self.provider == "ollama" and response_format is not None:
|
|
return await self._call_ollama_native(
|
|
messages,
|
|
response_format,
|
|
max_completion_tokens,
|
|
temperature,
|
|
max_retries,
|
|
initial_backoff,
|
|
max_backoff,
|
|
skip_validation,
|
|
start_time,
|
|
)
|
|
|
|
call_params = {
|
|
"model": self.model,
|
|
"messages": messages,
|
|
}
|
|
|
|
# Check if model supports reasoning parameter (o1, o3, gpt-5 families)
|
|
model_lower = self.model.lower()
|
|
is_reasoning_model = any(x in model_lower for x in ["gpt-5", "o1", "o3", "deepseek"])
|
|
|
|
# For GPT-4 and GPT-4.1 models, cap max_completion_tokens to 32000
|
|
# For GPT-4o models, cap to 16384
|
|
is_gpt4_model = any(x in model_lower for x in ["gpt-4.1", "gpt-4-"])
|
|
is_gpt4o_model = "gpt-4o" in model_lower
|
|
if max_completion_tokens is not None:
|
|
if is_gpt4o_model and max_completion_tokens > 16384:
|
|
max_completion_tokens = 16384
|
|
elif is_gpt4_model and max_completion_tokens > 32000:
|
|
max_completion_tokens = 32000
|
|
# For reasoning models, max_completion_tokens includes reasoning + output tokens
|
|
# Enforce minimum of 16000 to ensure enough space for both
|
|
if is_reasoning_model and max_completion_tokens < 16000:
|
|
max_completion_tokens = 16000
|
|
call_params["max_completion_tokens"] = max_completion_tokens
|
|
|
|
# GPT-5/o1/o3 family doesn't support custom temperature (only default 1)
|
|
if temperature is not None and not is_reasoning_model:
|
|
call_params["temperature"] = temperature
|
|
|
|
# Set reasoning_effort for reasoning models (OpenAI gpt-5, o1, o3)
|
|
if is_reasoning_model:
|
|
call_params["reasoning_effort"] = self.reasoning_effort
|
|
|
|
# Provider-specific parameters
|
|
if self.provider == "groq":
|
|
call_params["seed"] = DEFAULT_LLM_SEED
|
|
extra_body = {"service_tier": "auto"}
|
|
# Only add reasoning parameters for reasoning models
|
|
if is_reasoning_model:
|
|
extra_body["include_reasoning"] = False
|
|
call_params["extra_body"] = extra_body
|
|
|
|
last_exception = None
|
|
|
|
for attempt in range(max_retries + 1):
|
|
try:
|
|
if response_format is not None:
|
|
schema = None
|
|
if hasattr(response_format, "model_json_schema"):
|
|
schema = response_format.model_json_schema()
|
|
|
|
if strict_schema and schema is not None:
|
|
# Use OpenAI's strict JSON schema enforcement
|
|
# This guarantees all required fields are returned
|
|
call_params["response_format"] = {
|
|
"type": "json_schema",
|
|
"json_schema": {
|
|
"name": "response",
|
|
"strict": True,
|
|
"schema": schema,
|
|
},
|
|
}
|
|
else:
|
|
# Soft enforcement: add schema to prompt and use json_object mode
|
|
if schema is not None:
|
|
schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
|
|
|
|
if call_params["messages"] and call_params["messages"][0].get("role") == "system":
|
|
call_params["messages"][0]["content"] += schema_msg
|
|
elif call_params["messages"]:
|
|
call_params["messages"][0]["content"] = (
|
|
schema_msg + "\n\n" + call_params["messages"][0]["content"]
|
|
)
|
|
if self.provider not in ("lmstudio", "ollama"):
|
|
call_params["response_format"] = {"type": "json_object"}
|
|
|
|
logger.debug(f"Sending request to {self.provider}/{self.model} (timeout={self.timeout})")
|
|
response = await self._client.chat.completions.create(**call_params)
|
|
logger.debug(f"Received response from {self.provider}/{self.model}")
|
|
|
|
content = response.choices[0].message.content
|
|
|
|
# For local models, they may wrap JSON in markdown code blocks
|
|
if self.provider in ("lmstudio", "ollama"):
|
|
clean_content = content
|
|
if "```json" in content:
|
|
clean_content = content.split("```json")[1].split("```")[0].strip()
|
|
elif "```" in content:
|
|
clean_content = content.split("```")[1].split("```")[0].strip()
|
|
try:
|
|
json_data = json.loads(clean_content)
|
|
except json.JSONDecodeError:
|
|
# Fallback to parsing raw content
|
|
json_data = json.loads(content)
|
|
else:
|
|
# Log raw LLM response for debugging JSON parse issues
|
|
try:
|
|
json_data = json.loads(content)
|
|
except json.JSONDecodeError as json_err:
|
|
# Truncate content for logging (first 500 and last 200 chars)
|
|
content_preview = content[:500] if content else "<empty>"
|
|
if content and len(content) > 700:
|
|
content_preview = f"{content[:500]}...TRUNCATED...{content[-200:]}"
|
|
logger.warning(
|
|
f"JSON parse error from LLM response (attempt {attempt + 1}/{max_retries + 1}): {json_err}\n"
|
|
f" Model: {self.provider}/{self.model}\n"
|
|
f" Content length: {len(content) if content else 0} chars\n"
|
|
f" Content preview: {content_preview!r}\n"
|
|
f" Finish reason: {response.choices[0].finish_reason if response.choices else 'unknown'}"
|
|
)
|
|
# Retry on JSON parse errors - LLM may return valid JSON on next attempt
|
|
if attempt < max_retries:
|
|
backoff = min(initial_backoff * (2**attempt), max_backoff)
|
|
await asyncio.sleep(backoff)
|
|
last_exception = json_err
|
|
continue
|
|
else:
|
|
logger.error(f"JSON parse error after {max_retries + 1} attempts, giving up")
|
|
raise
|
|
|
|
if skip_validation:
|
|
result = json_data
|
|
else:
|
|
result = response_format.model_validate(json_data)
|
|
else:
|
|
response = await self._client.chat.completions.create(**call_params)
|
|
result = response.choices[0].message.content
|
|
|
|
# Log slow calls
|
|
duration = time.time() - start_time
|
|
usage = response.usage
|
|
if duration > 10.0:
|
|
ratio = max(1, usage.completion_tokens) / usage.prompt_tokens
|
|
cached_tokens = 0
|
|
if hasattr(usage, "prompt_tokens_details") and usage.prompt_tokens_details:
|
|
cached_tokens = getattr(usage.prompt_tokens_details, "cached_tokens", 0) or 0
|
|
cache_info = f", cached_tokens={cached_tokens}" if cached_tokens > 0 else ""
|
|
logger.info(
|
|
f"slow llm call: model={self.provider}/{self.model}, "
|
|
f"input_tokens={usage.prompt_tokens}, output_tokens={usage.completion_tokens}, "
|
|
f"total_tokens={usage.total_tokens}{cache_info}, time={duration:.3f}s, ratio out/in={ratio:.2f}"
|
|
)
|
|
|
|
return result
|
|
|
|
except LengthFinishReasonError as e:
|
|
logger.warning(f"LLM output exceeded token limits: {str(e)}")
|
|
raise OutputTooLongError(
|
|
"LLM output exceeded token limits. Input may need to be split into smaller chunks."
|
|
) from e
|
|
|
|
except APIConnectionError as e:
|
|
last_exception = e
|
|
if attempt < max_retries:
|
|
status_code = getattr(e, "status_code", None) or getattr(
|
|
getattr(e, "response", None), "status_code", None
|
|
)
|
|
logger.warning(
|
|
f"Connection error, retrying... (attempt {attempt + 1}/{max_retries + 1}) - status_code={status_code}, message={e}"
|
|
)
|
|
backoff = min(initial_backoff * (2**attempt), max_backoff)
|
|
await asyncio.sleep(backoff)
|
|
continue
|
|
else:
|
|
logger.error(f"Connection error after {max_retries + 1} attempts: {str(e)}")
|
|
raise
|
|
|
|
except APIStatusError as e:
|
|
# Fast fail only on 401 (unauthorized) and 403 (forbidden) - these won't recover with retries
|
|
if e.status_code in (401, 403):
|
|
logger.error(f"Auth error (HTTP {e.status_code}), not retrying: {str(e)}")
|
|
raise
|
|
|
|
last_exception = e
|
|
if attempt < max_retries:
|
|
backoff = min(initial_backoff * (2**attempt), max_backoff)
|
|
jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
|
|
sleep_time = backoff + jitter
|
|
await asyncio.sleep(sleep_time)
|
|
else:
|
|
logger.error(f"API error after {max_retries + 1} attempts: {str(e)}")
|
|
raise
|
|
|
|
except Exception as e:
|
|
logger.error(f"Unexpected error during LLM call: {type(e).__name__}: {str(e)}")
|
|
raise
|
|
|
|
if last_exception:
|
|
raise last_exception
|
|
raise RuntimeError("LLM call failed after all retries with no exception captured")
|
|
|
|
async def _call_anthropic(
|
|
self,
|
|
messages: list[dict[str, str]],
|
|
response_format: Any | None,
|
|
max_completion_tokens: int | None,
|
|
max_retries: int,
|
|
initial_backoff: float,
|
|
max_backoff: float,
|
|
skip_validation: bool,
|
|
start_time: float,
|
|
) -> Any:
|
|
"""Handle Anthropic-specific API calls."""
|
|
from anthropic import APIConnectionError, APIStatusError, RateLimitError
|
|
|
|
# Convert OpenAI-style messages to Anthropic format
|
|
system_prompt = None
|
|
anthropic_messages = []
|
|
|
|
for msg in messages:
|
|
role = msg.get("role", "user")
|
|
content = msg.get("content", "")
|
|
|
|
if role == "system":
|
|
if system_prompt:
|
|
system_prompt += "\n\n" + content
|
|
else:
|
|
system_prompt = content
|
|
else:
|
|
anthropic_messages.append({"role": role, "content": content})
|
|
|
|
# Add JSON schema instruction if response_format is provided
|
|
if response_format is not None and hasattr(response_format, "model_json_schema"):
|
|
schema = response_format.model_json_schema()
|
|
schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
|
|
if system_prompt:
|
|
system_prompt += schema_msg
|
|
else:
|
|
system_prompt = schema_msg
|
|
|
|
# Prepare parameters
|
|
call_params = {
|
|
"model": self.model,
|
|
"messages": anthropic_messages,
|
|
"max_tokens": max_completion_tokens if max_completion_tokens is not None else 4096,
|
|
}
|
|
|
|
if system_prompt:
|
|
call_params["system"] = system_prompt
|
|
|
|
last_exception = None
|
|
|
|
for attempt in range(max_retries + 1):
|
|
try:
|
|
response = await self._anthropic_client.messages.create(**call_params)
|
|
|
|
# Anthropic response content is a list of blocks
|
|
content = ""
|
|
for block in response.content:
|
|
if block.type == "text":
|
|
content += block.text
|
|
|
|
if response_format is not None:
|
|
# Models may wrap JSON in markdown code blocks
|
|
clean_content = content
|
|
if "```json" in content:
|
|
clean_content = content.split("```json")[1].split("```")[0].strip()
|
|
elif "```" in content:
|
|
clean_content = content.split("```")[1].split("```")[0].strip()
|
|
|
|
try:
|
|
json_data = json.loads(clean_content)
|
|
except json.JSONDecodeError:
|
|
# Fallback to parsing raw content if markdown stripping failed
|
|
json_data = json.loads(content)
|
|
|
|
if skip_validation:
|
|
result = json_data
|
|
else:
|
|
result = response_format.model_validate(json_data)
|
|
else:
|
|
result = content
|
|
|
|
# Log slow calls
|
|
duration = time.time() - start_time
|
|
if duration > 10.0:
|
|
input_tokens = response.usage.input_tokens
|
|
output_tokens = response.usage.output_tokens
|
|
logger.info(
|
|
f"slow llm call: model={self.provider}/{self.model}, "
|
|
f"input_tokens={input_tokens}, output_tokens={output_tokens}, "
|
|
f"time={duration:.3f}s"
|
|
)
|
|
|
|
return result
|
|
|
|
except json.JSONDecodeError as e:
|
|
last_exception = e
|
|
if attempt < max_retries:
|
|
logger.warning("Anthropic returned invalid JSON, retrying...")
|
|
backoff = min(initial_backoff * (2**attempt), max_backoff)
|
|
await asyncio.sleep(backoff)
|
|
continue
|
|
else:
|
|
logger.error(f"Anthropic returned invalid JSON after {max_retries + 1} attempts")
|
|
raise
|
|
|
|
except (APIConnectionError, RateLimitError, APIStatusError) as e:
|
|
# Fast fail on 401/403
|
|
if isinstance(e, APIStatusError) and e.status_code in (401, 403):
|
|
logger.error(f"Anthropic auth error (HTTP {e.status_code}), not retrying: {str(e)}")
|
|
raise
|
|
|
|
last_exception = e
|
|
if attempt < max_retries:
|
|
# Check if it's a rate limit or server error
|
|
should_retry = isinstance(e, (APIConnectionError, RateLimitError)) or (
|
|
isinstance(e, APIStatusError) and e.status_code >= 500
|
|
)
|
|
|
|
if should_retry:
|
|
backoff = min(initial_backoff * (2**attempt), max_backoff)
|
|
jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
|
|
await asyncio.sleep(backoff + jitter)
|
|
continue
|
|
|
|
logger.error(f"Anthropic API error after {max_retries + 1} attempts: {str(e)}")
|
|
raise
|
|
|
|
except Exception as e:
|
|
logger.error(f"Unexpected error during Anthropic call: {type(e).__name__}: {str(e)}")
|
|
raise
|
|
|
|
if last_exception:
|
|
raise last_exception
|
|
raise RuntimeError("Anthropic call failed after all retries")
|
|
|
|
async def _call_ollama_native(
|
|
self,
|
|
messages: list[dict[str, str]],
|
|
response_format: Any,
|
|
max_completion_tokens: int | None,
|
|
temperature: float | None,
|
|
max_retries: int,
|
|
initial_backoff: float,
|
|
max_backoff: float,
|
|
skip_validation: bool,
|
|
start_time: float,
|
|
) -> Any:
|
|
"""
|
|
Call Ollama using native API with JSON schema enforcement.
|
|
|
|
Ollama's native API supports passing a full JSON schema in the 'format' parameter,
|
|
which provides better structured output control than the OpenAI-compatible API.
|
|
"""
|
|
# Get the JSON schema from the Pydantic model
|
|
schema = response_format.model_json_schema() if hasattr(response_format, "model_json_schema") else None
|
|
|
|
# Build the base URL for Ollama's native API
|
|
# Default OpenAI-compatible URL is http://localhost:11434/v1
|
|
# Native API is at http://localhost:11434/api/chat
|
|
base_url = self.base_url or "http://localhost:11434/v1"
|
|
if base_url.endswith("/v1"):
|
|
native_url = base_url[:-3] + "/api/chat"
|
|
else:
|
|
native_url = base_url.rstrip("/") + "/api/chat"
|
|
|
|
# Build request payload
|
|
payload = {
|
|
"model": self.model,
|
|
"messages": messages,
|
|
"stream": False,
|
|
}
|
|
|
|
# Add schema as format parameter for structured output
|
|
if schema:
|
|
payload["format"] = schema
|
|
|
|
# Add optional parameters with optimized defaults for Ollama
|
|
# Benchmarking shows num_ctx=16384 + num_batch=512 is optimal
|
|
options = {
|
|
"num_ctx": 16384, # 16k context window for larger prompts
|
|
"num_batch": 512, # Optimal batch size for prompt processing
|
|
}
|
|
if max_completion_tokens:
|
|
options["num_predict"] = max_completion_tokens
|
|
if temperature is not None:
|
|
options["temperature"] = temperature
|
|
payload["options"] = options
|
|
|
|
last_exception = None
|
|
|
|
async with httpx.AsyncClient(timeout=300.0) as client:
|
|
for attempt in range(max_retries + 1):
|
|
try:
|
|
response = await client.post(native_url, json=payload)
|
|
response.raise_for_status()
|
|
|
|
result = response.json()
|
|
content = result.get("message", {}).get("content", "")
|
|
|
|
# Parse JSON response
|
|
try:
|
|
json_data = json.loads(content)
|
|
except json.JSONDecodeError as json_err:
|
|
content_preview = content[:500] if content else "<empty>"
|
|
if content and len(content) > 700:
|
|
content_preview = f"{content[:500]}...TRUNCATED...{content[-200:]}"
|
|
logger.warning(
|
|
f"Ollama JSON parse error (attempt {attempt + 1}/{max_retries + 1}): {json_err}\n"
|
|
f" Model: ollama/{self.model}\n"
|
|
f" Content length: {len(content) if content else 0} chars\n"
|
|
f" Content preview: {content_preview!r}"
|
|
)
|
|
if attempt < max_retries:
|
|
backoff = min(initial_backoff * (2**attempt), max_backoff)
|
|
await asyncio.sleep(backoff)
|
|
last_exception = json_err
|
|
continue
|
|
else:
|
|
raise
|
|
|
|
# Validate against Pydantic model or return raw JSON
|
|
if skip_validation:
|
|
return json_data
|
|
else:
|
|
return response_format.model_validate(json_data)
|
|
|
|
except httpx.HTTPStatusError as e:
|
|
last_exception = e
|
|
if attempt < max_retries:
|
|
logger.warning(
|
|
f"Ollama HTTP error (attempt {attempt + 1}/{max_retries + 1}): {e.response.status_code}"
|
|
)
|
|
backoff = min(initial_backoff * (2**attempt), max_backoff)
|
|
await asyncio.sleep(backoff)
|
|
continue
|
|
else:
|
|
logger.error(f"Ollama HTTP error after {max_retries + 1} attempts: {e}")
|
|
raise
|
|
|
|
except httpx.RequestError as e:
|
|
last_exception = e
|
|
if attempt < max_retries:
|
|
logger.warning(f"Ollama connection error (attempt {attempt + 1}/{max_retries + 1}): {e}")
|
|
backoff = min(initial_backoff * (2**attempt), max_backoff)
|
|
await asyncio.sleep(backoff)
|
|
continue
|
|
else:
|
|
logger.error(f"Ollama connection error after {max_retries + 1} attempts: {e}")
|
|
raise
|
|
|
|
except Exception as e:
|
|
logger.error(f"Unexpected error during Ollama call: {type(e).__name__}: {e}")
|
|
raise
|
|
|
|
if last_exception:
|
|
raise last_exception
|
|
raise RuntimeError("Ollama call failed after all retries")
|
|
|
|
async def _call_gemini(
|
|
self,
|
|
messages: list[dict[str, str]],
|
|
response_format: Any | None,
|
|
max_retries: int,
|
|
initial_backoff: float,
|
|
max_backoff: float,
|
|
skip_validation: bool,
|
|
start_time: float,
|
|
) -> Any:
|
|
"""Handle Gemini-specific API calls."""
|
|
# Convert OpenAI-style messages to Gemini format
|
|
system_instruction = None
|
|
gemini_contents = []
|
|
|
|
for msg in messages:
|
|
role = msg.get("role", "user")
|
|
content = msg.get("content", "")
|
|
|
|
if role == "system":
|
|
if system_instruction:
|
|
system_instruction += "\n\n" + content
|
|
else:
|
|
system_instruction = content
|
|
elif role == "assistant":
|
|
gemini_contents.append(genai_types.Content(role="model", parts=[genai_types.Part(text=content)]))
|
|
else:
|
|
gemini_contents.append(genai_types.Content(role="user", parts=[genai_types.Part(text=content)]))
|
|
|
|
# Add JSON schema instruction if response_format is provided
|
|
if response_format is not None and hasattr(response_format, "model_json_schema"):
|
|
schema = response_format.model_json_schema()
|
|
schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
|
|
if system_instruction:
|
|
system_instruction += schema_msg
|
|
else:
|
|
system_instruction = schema_msg
|
|
|
|
# Build generation config
|
|
config_kwargs = {}
|
|
if system_instruction:
|
|
config_kwargs["system_instruction"] = system_instruction
|
|
if response_format is not None:
|
|
config_kwargs["response_mime_type"] = "application/json"
|
|
config_kwargs["response_schema"] = response_format
|
|
|
|
generation_config = genai_types.GenerateContentConfig(**config_kwargs) if config_kwargs else None
|
|
|
|
last_exception = None
|
|
|
|
for attempt in range(max_retries + 1):
|
|
try:
|
|
response = await self._gemini_client.aio.models.generate_content(
|
|
model=self.model,
|
|
contents=gemini_contents,
|
|
config=generation_config,
|
|
)
|
|
|
|
content = response.text
|
|
|
|
# Handle empty response
|
|
if content is None:
|
|
block_reason = None
|
|
if hasattr(response, "candidates") and response.candidates:
|
|
candidate = response.candidates[0]
|
|
if hasattr(candidate, "finish_reason"):
|
|
block_reason = candidate.finish_reason
|
|
|
|
if attempt < max_retries:
|
|
logger.warning(f"Gemini returned empty response (reason: {block_reason}), retrying...")
|
|
backoff = min(initial_backoff * (2**attempt), max_backoff)
|
|
await asyncio.sleep(backoff)
|
|
continue
|
|
else:
|
|
raise RuntimeError(f"Gemini returned empty response after {max_retries + 1} attempts")
|
|
|
|
if response_format is not None:
|
|
json_data = json.loads(content)
|
|
if skip_validation:
|
|
result = json_data
|
|
else:
|
|
result = response_format.model_validate(json_data)
|
|
else:
|
|
result = content
|
|
|
|
# Log slow calls
|
|
duration = time.time() - start_time
|
|
if duration > 10.0 and hasattr(response, "usage_metadata") and response.usage_metadata:
|
|
usage = response.usage_metadata
|
|
logger.info(
|
|
f"slow llm call: model={self.provider}/{self.model}, "
|
|
f"input_tokens={usage.prompt_token_count}, output_tokens={usage.candidates_token_count}, "
|
|
f"time={duration:.3f}s"
|
|
)
|
|
|
|
return result
|
|
|
|
except json.JSONDecodeError as e:
|
|
last_exception = e
|
|
if attempt < max_retries:
|
|
logger.warning("Gemini returned invalid JSON, retrying...")
|
|
backoff = min(initial_backoff * (2**attempt), max_backoff)
|
|
await asyncio.sleep(backoff)
|
|
continue
|
|
else:
|
|
logger.error(f"Gemini returned invalid JSON after {max_retries + 1} attempts")
|
|
raise
|
|
|
|
except genai_errors.APIError as e:
|
|
# Fast fail only on 401 (unauthorized) and 403 (forbidden) - these won't recover with retries
|
|
if e.code in (401, 403):
|
|
logger.error(f"Gemini auth error (HTTP {e.code}), not retrying: {str(e)}")
|
|
raise
|
|
|
|
# Retry on retryable errors (rate limits, server errors, and other client errors like 400)
|
|
if e.code in (400, 429, 500, 502, 503, 504) or (e.code and e.code >= 500):
|
|
last_exception = e
|
|
if attempt < max_retries:
|
|
backoff = min(initial_backoff * (2**attempt), max_backoff)
|
|
jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
|
|
await asyncio.sleep(backoff + jitter)
|
|
else:
|
|
logger.error(f"Gemini API error after {max_retries + 1} attempts: {str(e)}")
|
|
raise
|
|
else:
|
|
logger.error(f"Gemini API error: {type(e).__name__}: {str(e)}")
|
|
raise
|
|
|
|
except Exception as e:
|
|
logger.error(f"Unexpected error during Gemini call: {type(e).__name__}: {str(e)}")
|
|
raise
|
|
|
|
if last_exception:
|
|
raise last_exception
|
|
raise RuntimeError("Gemini call failed after all retries")
|
|
|
|
@classmethod
|
|
def for_memory(cls) -> "LLMProvider":
|
|
"""Create provider for memory operations from environment variables."""
|
|
provider = os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq")
|
|
api_key = os.getenv("HINDSIGHT_API_LLM_API_KEY")
|
|
if not api_key:
|
|
raise ValueError("HINDSIGHT_API_LLM_API_KEY environment variable is required")
|
|
base_url = os.getenv("HINDSIGHT_API_LLM_BASE_URL", "")
|
|
model = os.getenv("HINDSIGHT_API_LLM_MODEL", "openai/gpt-oss-120b")
|
|
|
|
return cls(provider=provider, api_key=api_key, base_url=base_url, model=model, reasoning_effort="low")
|
|
|
|
@classmethod
|
|
def for_answer_generation(cls) -> "LLMProvider":
|
|
"""Create provider for answer generation. Falls back to memory config if not set."""
|
|
provider = os.getenv("HINDSIGHT_API_ANSWER_LLM_PROVIDER", os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq"))
|
|
api_key = os.getenv("HINDSIGHT_API_ANSWER_LLM_API_KEY", os.getenv("HINDSIGHT_API_LLM_API_KEY"))
|
|
if not api_key:
|
|
raise ValueError(
|
|
"HINDSIGHT_API_LLM_API_KEY or HINDSIGHT_API_ANSWER_LLM_API_KEY environment variable is required"
|
|
)
|
|
base_url = os.getenv("HINDSIGHT_API_ANSWER_LLM_BASE_URL", os.getenv("HINDSIGHT_API_LLM_BASE_URL", ""))
|
|
model = os.getenv("HINDSIGHT_API_ANSWER_LLM_MODEL", os.getenv("HINDSIGHT_API_LLM_MODEL", "openai/gpt-oss-120b"))
|
|
|
|
return cls(provider=provider, api_key=api_key, base_url=base_url, model=model, reasoning_effort="high")
|
|
|
|
@classmethod
|
|
def for_judge(cls) -> "LLMProvider":
|
|
"""Create provider for judge/evaluator operations. Falls back to memory config if not set."""
|
|
provider = os.getenv("HINDSIGHT_API_JUDGE_LLM_PROVIDER", os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq"))
|
|
api_key = os.getenv("HINDSIGHT_API_JUDGE_LLM_API_KEY", os.getenv("HINDSIGHT_API_LLM_API_KEY"))
|
|
if not api_key:
|
|
raise ValueError(
|
|
"HINDSIGHT_API_LLM_API_KEY or HINDSIGHT_API_JUDGE_LLM_API_KEY environment variable is required"
|
|
)
|
|
base_url = os.getenv("HINDSIGHT_API_JUDGE_LLM_BASE_URL", os.getenv("HINDSIGHT_API_LLM_BASE_URL", ""))
|
|
model = os.getenv("HINDSIGHT_API_JUDGE_LLM_MODEL", os.getenv("HINDSIGHT_API_LLM_MODEL", "openai/gpt-oss-120b"))
|
|
|
|
return cls(provider=provider, api_key=api_key, base_url=base_url, model=model, reasoning_effort="high")
|
|
|
|
|
|
# Backwards compatibility alias
|
|
LLMConfig = LLMProvider
|