86 lines
3.3 KiB
Python
86 lines
3.3 KiB
Python
import os
|
|
from typing import Any, Optional
|
|
|
|
from langchain_google_genai import ChatGoogleGenerativeAI
|
|
|
|
from .base_client import BaseLLMClient
|
|
from .validators import validate_model
|
|
|
|
|
|
class NormalizedChatGoogleGenerativeAI(ChatGoogleGenerativeAI):
|
|
"""ChatGoogleGenerativeAI with normalized content output.
|
|
|
|
Gemini 3 models return content as list: [{'type': 'text', 'text': '...'}]
|
|
This normalizes to string for consistent downstream handling.
|
|
"""
|
|
|
|
def _normalize_content(self, response):
|
|
content = response.content
|
|
if isinstance(content, list):
|
|
texts = [
|
|
item.get("text", "") if isinstance(item, dict) and item.get("type") == "text"
|
|
else item if isinstance(item, str) else ""
|
|
for item in content
|
|
]
|
|
response.content = "\n".join(t for t in texts if t)
|
|
return response
|
|
|
|
def invoke(self, input, config=None, **kwargs):
|
|
return self._normalize_content(super().invoke(input, config, **kwargs))
|
|
|
|
|
|
class GoogleClient(BaseLLMClient):
|
|
"""Client for Google Gemini models."""
|
|
|
|
def __init__(self, model: str, base_url: Optional[str] = None, **kwargs):
|
|
super().__init__(model, base_url, **kwargs)
|
|
|
|
def get_llm(self) -> Any:
|
|
"""Return configured ChatGoogleGenerativeAI instance."""
|
|
import certifi
|
|
|
|
# Fix SSL certificate path issue on Windows with conda
|
|
# Conda sets SSL_CERT_FILE to a non-existent path, so we clear it
|
|
# and let certifi handle it properly
|
|
ssl_cert_file = os.environ.get("SSL_CERT_FILE", "")
|
|
if ssl_cert_file and not os.path.exists(ssl_cert_file):
|
|
# Remove invalid SSL_CERT_FILE and use certifi instead
|
|
os.environ.pop("SSL_CERT_FILE", None)
|
|
os.environ["SSL_CERT_FILE"] = certifi.where()
|
|
|
|
llm_kwargs = {"model": self.model}
|
|
|
|
# Get Google API key from kwargs or environment
|
|
if "google_api_key" in self.kwargs:
|
|
llm_kwargs["google_api_key"] = self.kwargs["google_api_key"]
|
|
else:
|
|
api_key = os.environ.get("GOOGLE_API_KEY")
|
|
if api_key:
|
|
llm_kwargs["google_api_key"] = api_key
|
|
|
|
for key in ("timeout", "max_retries", "callbacks"):
|
|
if key in self.kwargs:
|
|
llm_kwargs[key] = self.kwargs[key]
|
|
|
|
# Map thinking_level to appropriate API param based on model
|
|
# Gemini 3 Pro: low, high
|
|
# Gemini 3 Flash: minimal, low, medium, high
|
|
# Gemini 2.5: thinking_budget (0=disable, -1=dynamic)
|
|
thinking_level = self.kwargs.get("thinking_level")
|
|
if thinking_level:
|
|
model_lower = self.model.lower()
|
|
if "gemini-3" in model_lower:
|
|
# Gemini 3 Pro doesn't support "minimal", use "low" instead
|
|
if "pro" in model_lower and thinking_level == "minimal":
|
|
thinking_level = "low"
|
|
llm_kwargs["thinking_level"] = thinking_level
|
|
else:
|
|
# Gemini 2.5: map to thinking_budget
|
|
llm_kwargs["thinking_budget"] = -1 if thinking_level == "high" else 0
|
|
|
|
return NormalizedChatGoogleGenerativeAI(**llm_kwargs)
|
|
|
|
def validate_model(self) -> bool:
|
|
"""Validate model for Google."""
|
|
return validate_model("google", self.model)
|