|
| 1 | +import base64 |
| 2 | +from collections.abc import Callable |
| 3 | +import os |
| 4 | +import time |
| 5 | +from typing import Any |
| 6 | + |
| 7 | +from google import genai |
| 8 | +from google.genai import types |
| 9 | +from google.genai.types import FunctionCall, FunctionCallingConfigMode, GenerateContentResponse |
| 10 | + |
| 11 | +from not_again_ai.llm.chat_completion.types import ( |
| 12 | + AssistantMessage, |
| 13 | + ChatCompletionChoice, |
| 14 | + ChatCompletionRequest, |
| 15 | + ChatCompletionResponse, |
| 16 | + Function, |
| 17 | + ImageContent, |
| 18 | + Role, |
| 19 | + TextContent, |
| 20 | + ToolCall, |
| 21 | +) |
| 22 | + |
| 23 | +# This should be all of the options we want to support in types.GenerateContentConfig, that are not handled otherwise |
| 24 | +GEMINI_PARAMETER_MAP = { |
| 25 | + "max_completion_tokens": "max_output_tokens", |
| 26 | + "temperature": "temperature", |
| 27 | + "top_p": "top_p", |
| 28 | + "top_k": "top_k", |
| 29 | +} |
| 30 | + |
| 31 | +GEMINI_FINISH_REASON_MAP = { |
| 32 | + "STOP": "stop", |
| 33 | + "MAX_TOKENS": "max_tokens", |
| 34 | + "SAFETY": "safety", |
| 35 | + "RECITATION": "recitation", |
| 36 | + "LANGUAGE": "language", |
| 37 | + "OTHER": "other", |
| 38 | + "BLOCKLIST": "blocklist", |
| 39 | + "PROHIBITED_CONTENT": "prohibited_content", |
| 40 | + "SPII": "spii", |
| 41 | + "MALFORMED_FUNCTION_CALL": "malformed_function_call", |
| 42 | + "IMAGE_SAFETY": "image_safety", |
| 43 | +} |
| 44 | + |
| 45 | + |
| 46 | +def gemini_chat_completion(request: ChatCompletionRequest, client: Callable[..., Any]) -> ChatCompletionResponse: |
| 47 | + """Experimental Gemini chat completion function.""" |
| 48 | + # Handle messages |
| 49 | + # Any system messages need to be removed from messages and concatenated into a single string (in order) |
| 50 | + system = "" |
| 51 | + contents = [] |
| 52 | + for message in request.messages: |
| 53 | + if message.role == "system": |
| 54 | + # Handle both string content and structured content |
| 55 | + if isinstance(message.content, str): |
| 56 | + system += message.content + "\n" |
| 57 | + else: |
| 58 | + # If it's a list of content parts, extract text content |
| 59 | + for part in message.content: |
| 60 | + if hasattr(part, "text"): |
| 61 | + system += part.text + "\n" |
| 62 | + elif message.role == "tool": |
| 63 | + tool_name = message.name if message.name is not None else "" |
| 64 | + function_response_part = types.Part.from_function_response( |
| 65 | + name=tool_name, |
| 66 | + response={"result": message.content}, |
| 67 | + ) |
| 68 | + contents.append( |
| 69 | + types.Content( |
| 70 | + role="user", |
| 71 | + parts=[function_response_part], |
| 72 | + ) |
| 73 | + ) |
| 74 | + elif message.role == "assistant": |
| 75 | + if message.content and isinstance(message.content, str): |
| 76 | + contents.append(types.Content(role="model", parts=[types.Part(text=message.content)])) |
| 77 | + function_parts = [] |
| 78 | + if isinstance(message, AssistantMessage) and message.tool_calls: |
| 79 | + for tool_call in message.tool_calls: |
| 80 | + function_call_part = types.Part( |
| 81 | + function_call=FunctionCall( |
| 82 | + id=tool_call.id, |
| 83 | + name=tool_call.function.name, |
| 84 | + args=tool_call.function.arguments, |
| 85 | + ) |
| 86 | + ) |
| 87 | + function_parts.append(function_call_part) |
| 88 | + if function_parts: |
| 89 | + contents.append(types.Content(role="model", parts=function_parts)) |
| 90 | + elif message.role == "user": |
| 91 | + if isinstance(message.content, str): |
| 92 | + contents.append(types.Content(role="user", parts=[types.Part(text=message.content)])) |
| 93 | + elif isinstance(message.content, list): |
| 94 | + parts = [] |
| 95 | + for part in message.content: |
| 96 | + if isinstance(part, TextContent): |
| 97 | + parts.append(types.Part(text=part.text)) |
| 98 | + elif isinstance(part, ImageContent): |
| 99 | + # Extract MIME type and data from data URI |
| 100 | + uri_parts = part.image_url.url.split(",", 1) |
| 101 | + if len(uri_parts) == 2: |
| 102 | + mime_type = uri_parts[0].split(":")[1].split(";")[0] |
| 103 | + base64_data = uri_parts[1] |
| 104 | + image_data = base64.b64decode(base64_data) |
| 105 | + parts.append(types.Part.from_bytes(mime_type=mime_type, data=image_data)) |
| 106 | + contents.append(types.Content(role="user", parts=parts)) |
| 107 | + |
| 108 | + kwargs: dict[str, Any] = {} |
| 109 | + kwargs["contents"] = contents |
| 110 | + kwargs["model"] = request.model |
| 111 | + config: dict[str, Any] = {} |
| 112 | + config["system_instruction"] = system.rstrip() |
| 113 | + config["automatic_function_calling"] = {"disable": True} |
| 114 | + |
| 115 | + # Handle the possible tool choice options |
| 116 | + if request.tool_choice: |
| 117 | + tool_choice = request.tool_choice |
| 118 | + if tool_choice == "auto": |
| 119 | + config["tool_config"] = types.FunctionCallingConfig(mode=FunctionCallingConfigMode.AUTO) |
| 120 | + elif tool_choice == "any": |
| 121 | + config["tool_config"] = types.FunctionCallingConfig(mode=FunctionCallingConfigMode.ANY) |
| 122 | + elif tool_choice == "none": |
| 123 | + config["tool_config"] = types.FunctionCallingConfig(mode=FunctionCallingConfigMode.NONE) |
| 124 | + elif isinstance(tool_choice, list): |
| 125 | + config["tool_config"] = types.FunctionCallingConfig( |
| 126 | + mode=FunctionCallingConfigMode.ANY, allowed_function_names=tool_choice |
| 127 | + ) |
| 128 | + elif tool_choice not in (None, "auto", "any", "none"): |
| 129 | + config["tool_config"] = types.FunctionCallingConfig( |
| 130 | + mode=FunctionCallingConfigMode.ANY, allowed_function_names=[tool_choice] |
| 131 | + ) |
| 132 | + |
| 133 | + # Handle tools |
| 134 | + tools = [] |
| 135 | + for tool in request.tools or []: |
| 136 | + tools.append(types.Tool(function_declarations=[tool])) # type: ignore |
| 137 | + if tools: |
| 138 | + config["tools"] = tools |
| 139 | + |
| 140 | + # Everything else defined in GEMINI_PARAMETER_MAP goes into kwargs["config"] |
| 141 | + request_kwargs = request.model_dump(mode="json", exclude_none=True) |
| 142 | + for key, value in GEMINI_PARAMETER_MAP.items(): |
| 143 | + if value is not None and key in request_kwargs: |
| 144 | + config[value] = request_kwargs.pop(key) |
| 145 | + |
| 146 | + kwargs["config"] = types.GenerateContentConfig(**config) |
| 147 | + |
| 148 | + start_time = time.time() |
| 149 | + response: GenerateContentResponse = client(**kwargs) |
| 150 | + end_time = time.time() |
| 151 | + response_duration = round(end_time - start_time, 4) |
| 152 | + |
| 153 | + finish_reason = "other" |
| 154 | + if response.candidates and response.candidates[0].finish_reason: |
| 155 | + finish_reason_str = str(response.candidates[0].finish_reason) |
| 156 | + finish_reason = GEMINI_FINISH_REASON_MAP.get(finish_reason_str, "other") |
| 157 | + |
| 158 | + tool_calls: list[ToolCall] = [] |
| 159 | + tool_call_objs = response.function_calls |
| 160 | + if tool_call_objs: |
| 161 | + for tool_call_obj in tool_call_objs: |
| 162 | + tool_call_id = tool_call_obj.id if tool_call_obj.id else "" |
| 163 | + tool_calls.append( |
| 164 | + ToolCall( |
| 165 | + id=tool_call_id, |
| 166 | + function=Function( |
| 167 | + name=tool_call_obj.name if tool_call_obj.name is not None else "", |
| 168 | + arguments=tool_call_obj.args if tool_call_obj.args is not None else {}, |
| 169 | + ), |
| 170 | + ) |
| 171 | + ) |
| 172 | + |
| 173 | + assistant_message = "" |
| 174 | + if ( |
| 175 | + response.candidates |
| 176 | + and response.candidates[0].content |
| 177 | + and response.candidates[0].content.parts |
| 178 | + and response.candidates[0].content.parts[0].text |
| 179 | + ): |
| 180 | + assistant_message = response.candidates[0].content.parts[0].text |
| 181 | + |
| 182 | + choice = ChatCompletionChoice( |
| 183 | + message=AssistantMessage( |
| 184 | + role=Role.ASSISTANT, |
| 185 | + content=assistant_message, |
| 186 | + tool_calls=tool_calls, |
| 187 | + ), |
| 188 | + finish_reason=finish_reason, |
| 189 | + ) |
| 190 | + |
| 191 | + completion_tokens = 0 |
| 192 | + # Add null check for usage_metadata |
| 193 | + if response.usage_metadata is not None: |
| 194 | + if response.usage_metadata.thoughts_token_count: |
| 195 | + completion_tokens = response.usage_metadata.thoughts_token_count |
| 196 | + if response.usage_metadata.candidates_token_count: |
| 197 | + completion_tokens += response.usage_metadata.candidates_token_count |
| 198 | + |
| 199 | + # Set safe default for prompt_tokens |
| 200 | + prompt_tokens = 0 |
| 201 | + if response.usage_metadata is not None and response.usage_metadata.prompt_token_count: |
| 202 | + prompt_tokens = response.usage_metadata.prompt_token_count |
| 203 | + |
| 204 | + chat_completion_response = ChatCompletionResponse( |
| 205 | + choices=[choice], |
| 206 | + completion_tokens=completion_tokens, |
| 207 | + prompt_tokens=prompt_tokens, |
| 208 | + response_duration=response_duration, |
| 209 | + ) |
| 210 | + return chat_completion_response |
| 211 | + |
| 212 | + |
| 213 | +def create_client_callable(client_class: type[genai.Client], **client_args: Any) -> Callable[..., Any]: |
| 214 | + """Creates a callable that instantiates and uses a Google genai client. |
| 215 | +
|
| 216 | + Args: |
| 217 | + client_class: The Google genai client class to instantiate |
| 218 | + **client_args: Arguments to pass to the client constructor |
| 219 | +
|
| 220 | + Returns: |
| 221 | + A callable that creates a client and returns completion results |
| 222 | + """ |
| 223 | + filtered_args = {k: v for k, v in client_args.items() if v is not None} |
| 224 | + |
| 225 | + def client_callable(**kwargs: Any) -> Any: |
| 226 | + client = client_class(**filtered_args) |
| 227 | + completion = client.models.generate_content(**kwargs) |
| 228 | + return completion |
| 229 | + |
| 230 | + return client_callable |
| 231 | + |
| 232 | + |
| 233 | +def gemini_client(api_key: str | None = None) -> Callable[..., Any]: |
| 234 | + if not api_key: |
| 235 | + api_key = os.environ.get("GEMINI_API_KEY") |
| 236 | + client_callable = create_client_callable(genai.Client, api_key=api_key) |
| 237 | + return client_callable |
0 commit comments