vertextoopenai / app /api_helpers.py
bibibi12345's picture
added extra safety filter option
a6fc424
import json
import time
import math
import asyncio
from typing import List, Dict, Any, Callable, Union, Optional
from fastapi.responses import JSONResponse, StreamingResponse
from google.auth.transport.requests import Request as AuthRequest
from google.genai import types
from openai import AsyncOpenAI
from models import OpenAIRequest, OpenAIMessage
from message_processing import (
convert_to_openai_format,
convert_chunk_to_openai,
extract_reasoning_by_tags,
_create_safety_ratings_html
)
import config as app_config
from config import VERTEX_REASONING_TAG
class StreamingReasoningProcessor:
def __init__(self, tag_name: str = VERTEX_REASONING_TAG):
self.tag_name = tag_name
self.open_tag = f"<{tag_name}>"
self.close_tag = f"</{tag_name}>"
self.tag_buffer = ""
self.inside_tag = False
self.reasoning_buffer = ""
self.partial_tag_buffer = ""
def process_chunk(self, content: str) -> tuple[str, str]:
if self.partial_tag_buffer:
content = self.partial_tag_buffer + content
self.partial_tag_buffer = ""
self.tag_buffer += content
processed_content = ""
current_reasoning = ""
while self.tag_buffer:
if not self.inside_tag:
open_pos = self.tag_buffer.find(self.open_tag)
if open_pos == -1:
partial_match = False
for i in range(1, min(len(self.open_tag), len(self.tag_buffer) + 1)):
if self.tag_buffer[-i:] == self.open_tag[:i]:
partial_match = True
if len(self.tag_buffer) > i:
processed_content += self.tag_buffer[:-i]
self.partial_tag_buffer = self.tag_buffer[-i:]
else: self.partial_tag_buffer = self.tag_buffer
self.tag_buffer = ""
break
if not partial_match:
processed_content += self.tag_buffer
self.tag_buffer = ""
break
else:
processed_content += self.tag_buffer[:open_pos]
self.tag_buffer = self.tag_buffer[open_pos + len(self.open_tag):]
self.inside_tag = True
else:
close_pos = self.tag_buffer.find(self.close_tag)
if close_pos == -1:
partial_match = False
for i in range(1, min(len(self.close_tag), len(self.tag_buffer) + 1)):
if self.tag_buffer[-i:] == self.close_tag[:i]:
partial_match = True
if len(self.tag_buffer) > i:
new_reasoning = self.tag_buffer[:-i]
self.reasoning_buffer += new_reasoning
if new_reasoning: current_reasoning = new_reasoning
self.partial_tag_buffer = self.tag_buffer[-i:]
else: self.partial_tag_buffer = self.tag_buffer
self.tag_buffer = ""
break
if not partial_match:
if self.tag_buffer:
self.reasoning_buffer += self.tag_buffer
current_reasoning = self.tag_buffer
self.tag_buffer = ""
break
else:
final_reasoning_chunk = self.tag_buffer[:close_pos]
self.reasoning_buffer += final_reasoning_chunk
if final_reasoning_chunk: current_reasoning = final_reasoning_chunk
self.reasoning_buffer = ""
self.tag_buffer = self.tag_buffer[close_pos + len(self.close_tag):]
self.inside_tag = False
return processed_content, current_reasoning
def flush_remaining(self) -> tuple[str, str]:
remaining_content, remaining_reasoning = "", ""
if self.partial_tag_buffer:
remaining_content += self.partial_tag_buffer
self.partial_tag_buffer = ""
if not self.inside_tag:
if self.tag_buffer: remaining_content += self.tag_buffer
else:
if self.reasoning_buffer: remaining_reasoning = self.reasoning_buffer
if self.tag_buffer: remaining_content += self.tag_buffer
self.inside_tag = False
self.tag_buffer, self.reasoning_buffer = "", ""
return remaining_content, remaining_reasoning
def create_openai_error_response(status_code: int, message: str, error_type: str) -> Dict[str, Any]:
return {"error": {"message": message, "type": error_type, "code": status_code, "param": None}}
def create_generation_config(request: OpenAIRequest) -> Dict[str, Any]:
config: Dict[str, Any] = {}
if request.temperature is not None: config["temperature"] = request.temperature
if request.max_tokens is not None: config["max_output_tokens"] = request.max_tokens
if request.top_p is not None: config["top_p"] = request.top_p
if request.top_k is not None: config["top_k"] = request.top_k
if request.stop is not None: config["stop_sequences"] = request.stop
if request.seed is not None: config["seed"] = request.seed
if request.n is not None: config["candidate_count"] = request.n
safety_threshold = "BLOCK_NONE" if app_config.SAFETY_SCORE else "BLOCK_ONLY_HIGH"
config["safety_settings"] = [
types.SafetySetting(category="HARM_CATEGORY_HATE_SPEECH", threshold=safety_threshold),
types.SafetySetting(category="HARM_CATEGORY_DANGEROUS_CONTENT", threshold=safety_threshold),
types.SafetySetting(category="HARM_CATEGORY_SEXUALLY_EXPLICIT", threshold=safety_threshold),
types.SafetySetting(category="HARM_CATEGORY_HARASSMENT", threshold=safety_threshold),
types.SafetySetting(category="HARM_CATEGORY_CIVIC_INTEGRITY", threshold=safety_threshold),
types.SafetySetting(category="HARM_CATEGORY_UNSPECIFIED", threshold=safety_threshold),
types.SafetySetting(category="HARM_CATEGORY_IMAGE_HATE", threshold=safety_threshold),
types.SafetySetting(category="HARM_CATEGORY_IMAGE_DANGEROUS_CONTENT", threshold=safety_threshold),
types.SafetySetting(category="HARM_CATEGORY_IMAGE_HARASSMENT", threshold=safety_threshold),
types.SafetySetting(category="HARM_CATEGORY_IMAGE_SEXUALLY_EXPLICIT", threshold=safety_threshold),
]
# config["thinking_config"] = {"include_thoughts": True}
# 1. Add tools (function declarations)
function_declarations = []
if request.tools:
for tool in request.tools:
if tool.get("type") == "function":
# func_def = tool.get("function")
func_def = tool
if func_def:
# Extract only the fields accepted by the Gemini API
declaration = {
"name": func_def.get("name"),
"description": func_def.get("description"),
}
# Get parameters and remove the $schema field if it exists
parameters = func_def.get("parameters")
if isinstance(parameters, dict) and "$schema" in parameters:
parameters = parameters.copy()
del parameters["$schema"]
if parameters is not None:
declaration["parameters"] = parameters
# Remove keys with None values to keep the payload clean
declaration = {k: v for k, v in declaration.items() if v is not None}
if declaration.get("name"): # Ensure name exists
function_declarations.append(declaration)
if function_declarations:
config["tools"] = [{"function_declarations": function_declarations}]
# 2. Add tool_config (based on tool_choice)
tool_config = None
if request.tool_choice:
choice = request.tool_choice
mode = None
allowed_functions = None
if isinstance(choice, str):
if choice == "none":
mode = "NONE"
elif choice == "auto":
mode = "AUTO"
elif isinstance(choice, dict) and choice.get("type") == "function":
func_name = choice.get("function", {}).get("name")
if func_name:
mode = "ANY" # 'ANY' mode is used to force a specific function call
allowed_functions = [func_name]
# If a valid mode was parsed, build the tool_config
if mode:
config_dict = {"mode": mode}
if allowed_functions:
config_dict["allowed_function_names"] = allowed_functions
tool_config = {"function_calling_config": config_dict}
if tool_config:
config["tool_config"] = tool_config
return config
def is_gemini_response_valid(response: Any) -> bool:
if response is None: return False
if hasattr(response, 'text') and isinstance(response.text, str) and response.text.strip(): return True
if hasattr(response, 'candidates') and response.candidates:
for cand in response.candidates:
if hasattr(cand, 'text') and isinstance(cand.text, str) and cand.text.strip(): return True
if hasattr(cand, 'content') and hasattr(cand.content, 'parts') and cand.content.parts:
for part in cand.content.parts:
if hasattr(part, 'function_call'): return True
if hasattr(part, 'text') and isinstance(getattr(part, 'text', None), str) and getattr(part, 'text', '').strip(): return True
return False
async def _chunk_openai_response_dict_for_sse(
openai_response_dict: Dict[str, Any],
response_id_override: Optional[str] = None,
model_name_override: Optional[str] = None
):
resp_id = response_id_override or openai_response_dict.get("id", f"chatcmpl-fakestream-{int(time.time())}")
model_name = model_name_override or openai_response_dict.get("model", "unknown")
created_time = openai_response_dict.get("created", int(time.time()))
choices = openai_response_dict.get("choices", [])
if not choices:
yield f"data: {json.dumps({'id': resp_id, 'object': 'chat.completion.chunk', 'created': created_time, 'model': model_name, 'choices': [{'index': 0, 'delta': {}, 'finish_reason': 'error'}]})}\n\n"
yield "data: [DONE]\n\n"
return
for choice_idx, choice in enumerate(choices):
message = choice.get("message", {})
final_finish_reason = choice.get("finish_reason", "stop")
if message.get("tool_calls"):
tool_calls_list = message.get("tool_calls", [])
for tc_item_idx, tool_call_item in enumerate(tool_calls_list):
delta_tc_start = {
"tool_calls": [{
"index": tc_item_idx,
"id": tool_call_item["id"],
"type": "function",
"function": {"name": tool_call_item["function"]["name"], "arguments": ""}
}]
}
yield f"data: {json.dumps({'id': resp_id, 'object': 'chat.completion.chunk', 'created': created_time, 'model': model_name, 'choices': [{'index': choice_idx, 'delta': delta_tc_start, 'finish_reason': None}]})}\n\n"
await asyncio.sleep(0.01)
delta_tc_args = {
"tool_calls": [{
"index": tc_item_idx,
"id": tool_call_item["id"],
"function": {"arguments": tool_call_item["function"]["arguments"]}
}]
}
yield f"data: {json.dumps({'id': resp_id, 'object': 'chat.completion.chunk', 'created': created_time, 'model': model_name, 'choices': [{'index': choice_idx, 'delta': delta_tc_args, 'finish_reason': None}]})}\n\n"
await asyncio.sleep(0.01)
elif message.get("content") is not None or message.get("reasoning_content") is not None :
reasoning_content = message.get("reasoning_content", "")
actual_content = message.get("content")
if reasoning_content:
delta_reasoning = {"reasoning_content": reasoning_content}
yield f"data: {json.dumps({'id': resp_id, 'object': 'chat.completion.chunk', 'created': created_time, 'model': model_name, 'choices': [{'index': choice_idx, 'delta': delta_reasoning, 'finish_reason': None}]})}\n\n"
if actual_content is not None: await asyncio.sleep(0.05)
content_to_chunk = actual_content if actual_content is not None else ""
if actual_content is not None:
chunk_size = max(1, math.ceil(len(content_to_chunk) / 10)) if content_to_chunk else 1
if not content_to_chunk and not reasoning_content :
yield f"data: {json.dumps({'id': resp_id, 'object': 'chat.completion.chunk', 'created': created_time, 'model': model_name, 'choices': [{'index': choice_idx, 'delta': {'content': ''}, 'finish_reason': None}]})}\n\n"
else:
for i in range(0, len(content_to_chunk), chunk_size):
yield f"data: {json.dumps({'id': resp_id, 'object': 'chat.completion.chunk', 'created': created_time, 'model': model_name, 'choices': [{'index': choice_idx, 'delta': {'content': content_to_chunk[i:i+chunk_size]}, 'finish_reason': None}]})}\n\n"
if len(content_to_chunk) > chunk_size: await asyncio.sleep(0.05)
yield f"data: {json.dumps({'id': resp_id, 'object': 'chat.completion.chunk', 'created': created_time, 'model': model_name, 'choices': [{'index': choice_idx, 'delta': {}, 'finish_reason': final_finish_reason}]})}\n\n"
yield "data: [DONE]\n\n"
async def gemini_fake_stream_generator(
gemini_client_instance: Any,
model_for_api_call: str,
prompt_for_api_call: List[types.Content],
gen_config_dict_for_api_call: Dict[str, Any],
request_obj: OpenAIRequest,
is_auto_attempt: bool
):
model_name_for_log = getattr(gemini_client_instance, 'model_name', 'unknown_gemini_model_object')
print(f"FAKE STREAMING (Gemini): Prep for '{request_obj.model}' (API model string: '{model_for_api_call}', client obj: '{model_name_for_log}')")
api_call_task = asyncio.create_task(
gemini_client_instance.aio.models.generate_content(
model=model_for_api_call,
contents=prompt_for_api_call,
config=gen_config_dict_for_api_call # Pass the dictionary directly
)
)
outer_keep_alive_interval = app_config.FAKE_STREAMING_INTERVAL_SECONDS
if outer_keep_alive_interval > 0:
while not api_call_task.done():
keep_alive_data = {"id": "chatcmpl-keepalive", "object": "chat.completion.chunk", "created": int(time.time()), "model": request_obj.model, "choices": [{"delta": {"content": ""}, "index": 0, "finish_reason": None}]}
yield f"data: {json.dumps(keep_alive_data)}\n\n"
await asyncio.sleep(outer_keep_alive_interval)
try:
raw_gemini_response = await api_call_task
openai_response_dict = convert_to_openai_format(raw_gemini_response, request_obj.model)
if hasattr(raw_gemini_response, 'prompt_feedback') and \
hasattr(raw_gemini_response.prompt_feedback, 'block_reason') and \
raw_gemini_response.prompt_feedback.block_reason:
block_message = f"Response blocked by Gemini safety filter: {raw_gemini_response.prompt_feedback.block_reason}"
if hasattr(raw_gemini_response.prompt_feedback, 'block_reason_message') and \
raw_gemini_response.prompt_feedback.block_reason_message:
block_message += f" (Message: {raw_gemini_response.prompt_feedback.block_reason_message})"
raise ValueError(block_message)
async for chunk_sse in _chunk_openai_response_dict_for_sse(
openai_response_dict=openai_response_dict
):
yield chunk_sse
except Exception as e_outer_gemini:
err_msg_detail = f"Error in gemini_fake_stream_generator (model: '{request_obj.model}'): {type(e_outer_gemini).__name__} - {str(e_outer_gemini)}"
print(f"ERROR: {err_msg_detail}")
sse_err_msg_display = str(e_outer_gemini)
if len(sse_err_msg_display) > 512: sse_err_msg_display = sse_err_msg_display[:512] + "..."
err_resp_sse = create_openai_error_response(500, sse_err_msg_display, "server_error")
json_payload_error = json.dumps(err_resp_sse)
if not is_auto_attempt:
yield f"data: {json_payload_error}\n\n"
yield "data: [DONE]\n\n"
if is_auto_attempt: raise
async def openai_fake_stream_generator(
openai_client: Union[AsyncOpenAI, Any],
openai_params: Dict[str, Any],
openai_extra_body: Dict[str, Any],
request_obj: OpenAIRequest,
is_auto_attempt: bool
):
api_model_name = openai_params.get("model", "unknown-openai-model")
print(f"FAKE STREAMING (OpenAI Direct): Prep for '{request_obj.model}' (API model: '{api_model_name}')")
response_id = f"chatcmpl-openaidirectfake-{int(time.time())}"
async def _openai_api_call_task():
params_for_call = openai_params.copy()
params_for_call['stream'] = False
return await openai_client.chat.completions.create(**params_for_call, extra_body=openai_extra_body)
api_call_task = asyncio.create_task(_openai_api_call_task())
outer_keep_alive_interval = app_config.FAKE_STREAMING_INTERVAL_SECONDS
if outer_keep_alive_interval > 0:
while not api_call_task.done():
keep_alive_data = {"id": "chatcmpl-keepalive", "object": "chat.completion.chunk", "created": int(time.time()), "model": request_obj.model, "choices": [{"delta": {"content": ""}, "index": 0, "finish_reason": None}]}
yield f"data: {json.dumps(keep_alive_data)}\n\n"
await asyncio.sleep(outer_keep_alive_interval)
try:
raw_response_obj = await api_call_task
openai_response_dict = raw_response_obj.model_dump(exclude_unset=True, exclude_none=True)
if app_config.SAFETY_SCORE and hasattr(raw_response_obj, "choices") and raw_response_obj.choices:
for i, choice_obj in enumerate(raw_response_obj.choices):
if hasattr(choice_obj, "safety_ratings") and choice_obj.safety_ratings:
safety_html = _create_safety_ratings_html(choice_obj.safety_ratings)
if i < len(openai_response_dict.get("choices", [])):
choice_dict = openai_response_dict["choices"][i]
message_dict = choice_dict.get("message")
if message_dict:
current_content = message_dict.get("content") or ""
message_dict["content"] = current_content + safety_html
if openai_response_dict.get("choices") and \
isinstance(openai_response_dict["choices"], list) and \
len(openai_response_dict["choices"]) > 0:
first_choice_dict_item = openai_response_dict["choices"]
if first_choice_dict_item and isinstance(first_choice_dict_item, dict) :
choice_message_ref = first_choice_dict_item.get("message", {})
original_content = choice_message_ref.get("content")
if isinstance(original_content, str):
reasoning_text, actual_content = extract_reasoning_by_tags(original_content, VERTEX_REASONING_TAG)
choice_message_ref["content"] = actual_content
if reasoning_text:
choice_message_ref["reasoning_content"] = reasoning_text
async for chunk_sse in _chunk_openai_response_dict_for_sse(
openai_response_dict=openai_response_dict,
response_id_override=response_id,
model_name_override=request_obj.model
):
yield chunk_sse
except Exception as e_outer:
err_msg_detail = f"Error in openai_fake_stream_generator (model: '{request_obj.model}'): {type(e_outer).__name__} - {str(e_outer)}"
print(f"ERROR: {err_msg_detail}")
sse_err_msg_display = str(e_outer)
if len(sse_err_msg_display) > 512: sse_err_msg_display = sse_err_msg_display[:512] + "..."
err_resp_sse = create_openai_error_response(500, sse_err_msg_display, "server_error")
json_payload_error = json.dumps(err_resp_sse)
if not is_auto_attempt:
yield f"data: {json_payload_error}\n\n"
yield "data: [DONE]\n\n"
if is_auto_attempt: raise
async def execute_gemini_call(
current_client: Any,
model_to_call: str,
prompt_func: Callable[[List[OpenAIMessage]], List[types.Content]],
gen_config_dict: Dict[str, Any],
request_obj: OpenAIRequest,
is_auto_attempt: bool = False
):
actual_prompt_for_call = prompt_func(request_obj.messages)
client_model_name_for_log = getattr(current_client, 'model_name', 'unknown_direct_client_object')
print(f"INFO: execute_gemini_call for requested API model '{model_to_call}', using client object with internal name '{client_model_name_for_log}'. Original request model: '{request_obj.model}'")
if request_obj.stream:
if app_config.FAKE_STREAMING_ENABLED:
return StreamingResponse(
gemini_fake_stream_generator(
current_client, model_to_call, actual_prompt_for_call,
gen_config_dict,
request_obj, is_auto_attempt
), media_type="text/event-stream"
)
else: # True Streaming
response_id_for_stream = f"chatcmpl-realstream-{int(time.time())}"
async def _gemini_real_stream_generator_inner():
try:
stream_gen_obj = await current_client.aio.models.generate_content_stream(
model=model_to_call,
contents=actual_prompt_for_call,
config=gen_config_dict # Pass the dictionary directly
)
async for chunk_item_call in stream_gen_obj:
yield convert_chunk_to_openai(chunk_item_call, request_obj.model, response_id_for_stream, 0)
yield "data: [DONE]\n\n"
except Exception as e_stream_call:
err_msg_detail_stream = f"Streaming Error (Gemini API, model string: '{model_to_call}'): {type(e_stream_call).__name__} - {str(e_stream_call)}"
print(f"ERROR: {err_msg_detail_stream}")
s_err = str(e_stream_call); s_err = s_err[:1024]+"..." if len(s_err)>1024 else s_err
err_resp = create_openai_error_response(500,s_err,"server_error")
j_err = json.dumps(err_resp)
if not is_auto_attempt:
yield f"data: {j_err}\n\n"
yield "data: [DONE]\n\n"
raise e_stream_call
return StreamingResponse(_gemini_real_stream_generator_inner(), media_type="text/event-stream")
else: # Non-streaming
response_obj_call = await current_client.aio.models.generate_content(
model=model_to_call,
contents=actual_prompt_for_call,
config=gen_config_dict # Pass the dictionary directly
)
if hasattr(response_obj_call, 'prompt_feedback') and \
hasattr(response_obj_call.prompt_feedback, 'block_reason') and \
response_obj_call.prompt_feedback.block_reason:
block_msg = f"Blocked (Gemini): {response_obj_call.prompt_feedback.block_reason}"
if hasattr(response_obj_call.prompt_feedback,'block_reason_message') and \
response_obj_call.prompt_feedback.block_reason_message:
block_msg+=f" ({response_obj_call.prompt_feedback.block_reason_message})"
raise ValueError(block_msg)
if not is_gemini_response_valid(response_obj_call):
error_details = f"Invalid non-streaming Gemini response for model string '{model_to_call}'. "
if hasattr(response_obj_call, 'candidates'):
error_details += f"Candidates: {len(response_obj_call.candidates) if response_obj_call.candidates else 0}. "
if response_obj_call.candidates and len(response_obj_call.candidates) > 0:
candidate = response_obj_call.candidates if isinstance(response_obj_call.candidates, list) else response_obj_call.candidates
if hasattr(candidate, 'content'):
error_details += "Has content. "
if hasattr(candidate.content, 'parts'):
error_details += f"Parts: {len(candidate.content.parts) if candidate.content.parts else 0}. "
if candidate.content.parts and len(candidate.content.parts) > 0:
part = candidate.content.parts if isinstance(candidate.content.parts, list) else candidate.content.parts
if hasattr(part, 'text'):
text_preview = str(getattr(part, 'text', ''))[:100]
error_details += f"First part text: '{text_preview}'"
elif hasattr(part, 'function_call'):
error_details += f"First part is function_call: {part.function_call.name}"
else:
error_details += f"Response type: {type(response_obj_call).__name__}"
raise ValueError(error_details)
openai_response_content = convert_to_openai_format(response_obj_call, request_obj.model)
return JSONResponse(content=openai_response_content)