48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263 | class GeminiMultiModal(MultiModalLLM):
"""Gemini multimodal."""
model_name: str = Field(
default=GEMINI_MM_MODELS[0], description="The Gemini model to use."
)
temperature: float = Field(
default=DEFAULT_TEMPERATURE,
description="The temperature to use during generation.",
gte=0.0,
lte=1.0,
)
max_tokens: int = Field(
default=DEFAULT_NUM_OUTPUTS,
description="The number of tokens to generate.",
gt=0,
)
generate_kwargs: dict = Field(
default_factory=dict, description="Kwargs for generation."
)
_model: "genai.GenerativeModel" = PrivateAttr()
_model_meta: "genai.types.Model" = PrivateAttr()
def __init__(
self,
api_key: Optional[str] = None,
model_name: Optional[str] = GEMINI_MM_MODELS[0],
temperature: float = DEFAULT_TEMPERATURE,
max_tokens: Optional[int] = None,
generation_config: Optional["genai.types.GenerationConfigDict"] = None,
safety_settings: "genai.types.SafetySettingOptions" = None,
api_base: Optional[str] = None,
default_headers: Optional[Dict[str, str]] = None,
transport: Optional[str] = None,
callback_manager: Optional[CallbackManager] = None,
**generate_kwargs: Any,
):
"""Creates a new Gemini model interface."""
# API keys are optional. The API can be authorised via OAuth (detected
# environmentally) or by the GOOGLE_API_KEY environment variable.
config_params: Dict[str, Any] = {
"api_key": api_key or os.getenv("GOOGLE_API_KEY"),
}
if api_base:
config_params["client_options"] = {"api_endpoint": api_base}
if default_headers:
default_metadata: Sequence[Tuple[str, str]] = []
for key, value in default_headers.items():
default_metadata.append((key, value))
# `default_metadata` contains (key, value) pairs that will be sent with every request.
# These will be sent as HTTP headers When using `transport="rest"`.
config_params["default_metadata"] = default_metadata
if transport:
config_params["transport"] = transport
# transport: A string, one of: [`rest`, `grpc`, `grpc_asyncio`].
genai.configure(**config_params)
base_gen_config = generation_config if generation_config else {}
# Explicitly passed args take precedence over the generation_config.
final_gen_config = {"temperature": temperature} | base_gen_config
# Check whether the Gemini Model is supported or not
if model_name not in GEMINI_MM_MODELS:
raise ValueError(
f"Invalid model {model_name}. "
f"Available models are: {GEMINI_MM_MODELS}"
)
model = genai.GenerativeModel(
model_name=model_name,
generation_config=final_gen_config,
safety_settings=safety_settings,
)
model_meta = genai.get_model(model_name)
supported_methods = model_meta.supported_generation_methods
if "generateContent" not in supported_methods:
raise ValueError(
f"Model {model_name} does not support content generation, only "
f"{supported_methods}."
)
if not max_tokens:
max_tokens = model_meta.output_token_limit
else:
max_tokens = min(max_tokens, model_meta.output_token_limit)
super().__init__(
model_name=model_name,
temperature=temperature,
max_tokens=max_tokens,
generate_kwargs=generate_kwargs,
callback_manager=callback_manager,
)
self._model = model
self._model_meta = model_meta
@classmethod
def class_name(cls) -> str:
return "Gemini_MultiModal_LLM"
@property
def metadata(self) -> MultiModalLLMMetadata:
total_tokens = self._model_meta.input_token_limit + self.max_tokens
return MultiModalLLMMetadata(
context_window=total_tokens,
num_output=self.max_tokens,
model_name=self.model_name,
)
def complete(
self, prompt: str, image_documents: Sequence[ImageNode], **kwargs: Any
) -> CompletionResponse:
images = [PIL.Image.open(doc.resolve_image()) for doc in image_documents]
result = self._model.generate_content([prompt, *images], **kwargs)
return completion_from_gemini_response(result)
def stream_complete(
self, prompt: str, image_documents: Sequence[ImageNode], **kwargs: Any
) -> CompletionResponseGen:
images = [PIL.Image.open(doc.resolve_image()) for doc in image_documents]
result = self._model.generate_content([prompt, *images], stream=True, **kwargs)
yield from map(completion_from_gemini_response, result)
def chat(self, messages: Sequence[ChatMessage], **kwargs: Any) -> ChatResponse:
*history, next_msg = map(chat_message_to_gemini, messages)
chat = self._model.start_chat(history=history)
response = chat.send_message(next_msg)
return chat_from_gemini_response(response)
def stream_chat(
self, messages: Sequence[ChatMessage], **kwargs: Any
) -> ChatResponseGen:
*history, next_msg = map(chat_message_to_gemini, messages)
chat = self._model.start_chat(history=history)
response = chat.send_message(next_msg, stream=True)
def gen() -> ChatResponseGen:
content = ""
for r in response:
top_candidate = r.candidates[0]
content_delta = top_candidate.content.parts[0].text
role = ROLES_FROM_GEMINI[top_candidate.content.role]
raw = {
**(type(top_candidate).to_dict(top_candidate)),
**(
type(response.prompt_feedback).to_dict(response.prompt_feedback)
),
}
content += content_delta
yield ChatResponse(
message=ChatMessage(role=role, content=content),
delta=content_delta,
raw=raw,
)
return gen()
async def acomplete(
self, prompt: str, image_documents: Sequence[ImageNode], **kwargs: Any
) -> CompletionResponse:
images = [PIL.Image.open(doc.resolve_image()) for doc in image_documents]
result = await self._model.generate_content_async([prompt, *images], **kwargs)
return completion_from_gemini_response(result)
async def astream_complete(
self, prompt: str, image_documents: Sequence[ImageNode], **kwargs: Any
) -> CompletionResponseAsyncGen:
images = [PIL.Image.open(doc.resolve_image()) for doc in image_documents]
ait = await self._model.generate_content_async(
[prompt, *images], stream=True, **kwargs
)
async def gen() -> CompletionResponseAsyncGen:
async for comp in ait:
yield completion_from_gemini_response(comp)
return gen()
async def achat(
self, messages: Sequence[ChatMessage], **kwargs: Any
) -> ChatResponse:
*history, next_msg = map(chat_message_to_gemini, messages)
chat = self._model.start_chat(history=history)
response = await chat.send_message_async(next_msg)
return chat_from_gemini_response(response)
async def astream_chat(
self, messages: Sequence[ChatMessage], **kwargs: Any
) -> ChatResponseAsyncGen:
*history, next_msg = map(chat_message_to_gemini, messages)
chat = self._model.start_chat(history=history)
response = await chat.send_message_async(next_msg, stream=True)
async def gen() -> ChatResponseAsyncGen:
content = ""
for r in response:
top_candidate = r.candidates[0]
content_delta = top_candidate.content.parts[0].text
role = ROLES_FROM_GEMINI[top_candidate.content.role]
raw = {
**(type(top_candidate).to_dict(top_candidate)),
**(
type(response.prompt_feedback).to_dict(response.prompt_feedback)
),
}
content += content_delta
yield ChatResponse(
message=ChatMessage(role=role, content=content),
delta=content_delta,
raw=raw,
)
return gen()
|