31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
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 | @deprecated(
reason="This class has been deprecated and will no longer be maintained. Please use BedrockConverse from llama-index-llms-bedrock-converse instead. See Multi Modal LLMs documentation for a complete guide on migration: https://docs.llamaindex.ai/en/stable/understanding/using_llms/using_llms/#multi-modal-llms",
version="0.1.1",
)
class BedrockMultiModal(BedrockConverse):
"""Bedrock Multi-Modal LLM implementation."""
model: str = Field(description="The Multi-Modal model to use from Bedrock.")
temperature: float = Field(description="The temperature to use for sampling.")
max_tokens: Optional[int] = Field(
description="The maximum numbers of tokens to generate.",
gt=0,
)
context_window: Optional[int] = Field(
description="The maximum number of context tokens for the model.",
gt=0,
)
region_name: str = Field(
default=None,
description="AWS region name.",
)
aws_access_key_id: str = Field(
default=None,
description="AWS access key ID.",
exclude=True,
)
aws_secret_access_key: str = Field(
default=None,
description="AWS secret access key.",
exclude=True,
)
max_retries: int = Field(
default=10,
description="The maximum number of API retries.",
gt=0,
)
timeout: float = Field(
default=60.0,
description="The timeout for API requests in seconds.",
gt=0,
)
additional_kwargs: Dict[str, Any] = Field(
default_factory=dict,
description="Additional kwargs for the Bedrock API.",
)
_messages_to_prompt: Callable = PrivateAttr()
_completion_to_prompt: Callable = PrivateAttr()
_client: Any = PrivateAttr() # boto3 client
_config: Any = PrivateAttr() # botocore config
_asession: Any = PrivateAttr() # aioboto3 session
def __init__(
self,
model: str = "anthropic.claude-3-sonnet-20240229-v1:0",
temperature: float = DEFAULT_TEMPERATURE,
max_tokens: Optional[int] = 300,
additional_kwargs: Optional[Dict[str, Any]] = None,
context_window: Optional[int] = DEFAULT_CONTEXT_WINDOW,
region_name: Optional[str] = None,
aws_access_key_id: Optional[str] = None,
aws_secret_access_key: Optional[str] = None,
max_retries: int = 10,
timeout: float = 60.0,
messages_to_prompt: Optional[Callable] = None,
completion_to_prompt: Optional[Callable] = None,
callback_manager: Optional[CallbackManager] = None,
**kwargs: Any,
) -> None:
"""Initialize params."""
# Validate model name first
if model not in BEDROCK_MULTI_MODAL_MODELS:
raise ValueError(
f"Invalid model {model}. "
f"Available models are: {list(BEDROCK_MULTI_MODAL_MODELS.keys())}"
)
aws_access_key_id, aws_secret_access_key, region = resolve_bedrock_credentials(
region_name=region_name,
aws_access_key_id=aws_access_key_id,
aws_secret_access_key=aws_secret_access_key,
)
super().__init__(
model=model,
temperature=temperature,
max_tokens=max_tokens,
additional_kwargs=additional_kwargs or {},
context_window=context_window,
region_name=region,
aws_access_key_id=aws_access_key_id,
aws_secret_access_key=aws_secret_access_key,
max_retries=max_retries,
timeout=timeout,
callback_manager=callback_manager,
**kwargs,
)
self._messages_to_prompt = messages_to_prompt or generic_messages_to_prompt
self._completion_to_prompt = completion_to_prompt or (lambda x: x)
self._config = Config(
retries={"max_attempts": max_retries, "mode": "standard"},
connect_timeout=timeout,
read_timeout=timeout,
)
self._client = self._get_client()
self._asession = aioboto3.Session(
aws_access_key_id=self.aws_access_key_id,
aws_secret_access_key=self.aws_secret_access_key,
region_name=self.region_name,
)
@classmethod
def class_name(cls) -> str:
"""Get class name."""
return "bedrock_multi_modal_llm"
@property
def metadata(self) -> MultiModalLLMMetadata:
"""Multi Modal LLM metadata."""
return MultiModalLLMMetadata(
num_output=self.max_tokens or DEFAULT_NUM_OUTPUTS,
model_name=self.model,
)
def _get_model_kwargs(self, **kwargs: Any) -> Dict[str, Any]:
"""Get model kwargs."""
# For Claude models, parameters need to be part of the body
model_kwargs = {
"contentType": "application/json",
"accept": "application/json",
}
if self.model.startswith("anthropic.claude"):
model_kwargs["body"] = {
"anthropic_version": "bedrock-2023-05-31",
"max_tokens": self.max_tokens if self.max_tokens is not None else 300,
"temperature": self.temperature,
}
# Add any additional kwargs
if "body" in model_kwargs:
model_kwargs["body"].update(self.additional_kwargs)
model_kwargs["body"].update(kwargs)
return model_kwargs
def _complete(
self, prompt: str, image_documents: Sequence[ImageNode], **kwargs: Any
) -> CompletionResponse:
"""Complete the prompt with image support."""
message = generate_bedrock_multi_modal_message(
prompt=prompt,
image_documents=image_documents,
)
# Get model kwargs and prepare the request body
model_kwargs = self._get_model_kwargs(**kwargs)
response = super().chat(
messages=message**model_kwargs,
)
return CompletionResponse(
text=response.message.content or "",
raw=response.raw,
)
def complete(
self, prompt: str, image_documents: Sequence[ImageNode], **kwargs: Any
) -> CompletionResponse:
"""Complete the prompt with image support."""
return self._complete(prompt, image_documents, **kwargs)
async def acomplete(
self, prompt: str, image_documents: Sequence[ImageNode], **kwargs: Any
) -> CompletionResponse:
"""Complete the prompt with image support asynchronously."""
message = generate_bedrock_multi_modal_message(
prompt=prompt,
image_documents=image_documents,
)
# Get model kwargs and prepare the request body
model_kwargs = self._get_model_kwargs(**kwargs)
response = await super().achat(
messages=message**model_kwargs,
)
return CompletionResponse(
text=response.message.content or "",
raw=response.raw,
)
|