|
| 1 | +import os |
| 2 | +from typing import Any |
| 3 | +from typing import Callable |
| 4 | +from typing import Optional |
| 5 | +from typing import Union |
| 6 | + |
| 7 | +import httpx |
| 8 | + |
| 9 | +from singlestoredb import manage_workspaces |
| 10 | +from singlestoredb.management.inference_api import InferenceAPIInfo |
| 11 | + |
| 12 | +try: |
| 13 | + from langchain_openai import ChatOpenAI |
| 14 | +except ImportError: |
| 15 | + raise ImportError( |
| 16 | + 'Could not import langchain_openai python package. ' |
| 17 | + 'Please install it with `pip install langchain_openai`.', |
| 18 | + ) |
| 19 | + |
| 20 | +try: |
| 21 | + from langchain_aws import ChatBedrockConverse |
| 22 | +except ImportError: |
| 23 | + raise ImportError( |
| 24 | + 'Could not import langchain-aws python package. ' |
| 25 | + 'Please install it with `pip install langchain-aws`.', |
| 26 | + ) |
| 27 | + |
| 28 | +import boto3 |
| 29 | +from botocore import UNSIGNED |
| 30 | +from botocore.config import Config |
| 31 | + |
| 32 | + |
| 33 | +def SingleStoreChatFactoryDebug( |
| 34 | + model_name: str, |
| 35 | + api_key: Optional[Union[Optional[str], Callable[[], Optional[str]]]] = None, |
| 36 | + streaming: bool = True, |
| 37 | + http_client: Optional[httpx.Client] = None, |
| 38 | + obo_token: Optional[Union[Optional[str], Callable[[], Optional[str]]]] = None, |
| 39 | + obo_token_getter: Optional[Callable[[], Optional[str]]] = None, |
| 40 | + base_url: Optional[str] = None, |
| 41 | + hosting_platform: Optional[str] = None, |
| 42 | + timeout: Optional[float] = None, |
| 43 | + **kwargs: Any, |
| 44 | +) -> Union[ChatOpenAI, ChatBedrockConverse]: |
| 45 | + """Return a chat model instance (ChatOpenAI or ChatBedrockConverse). |
| 46 | + """ |
| 47 | + # Handle api_key and obo_token as callable functions |
| 48 | + if callable(api_key): |
| 49 | + api_key_getter_fn = api_key |
| 50 | + else: |
| 51 | + def api_key_getter_fn() -> Optional[str]: |
| 52 | + if api_key is None: |
| 53 | + return os.environ.get('SINGLESTOREDB_USER_TOKEN') |
| 54 | + return api_key |
| 55 | + |
| 56 | + if obo_token_getter is not None: |
| 57 | + obo_token_getter_fn = obo_token_getter |
| 58 | + else: |
| 59 | + if callable(obo_token): |
| 60 | + obo_token_getter_fn = obo_token |
| 61 | + else: |
| 62 | + def obo_token_getter_fn() -> Optional[str]: |
| 63 | + return obo_token |
| 64 | + |
| 65 | + # handle model info |
| 66 | + if base_url is None: |
| 67 | + base_url = os.environ.get('SINGLESTOREDB_INFERENCE_API_BASE_URL') |
| 68 | + if hosting_platform is None: |
| 69 | + hosting_platform = os.environ.get('SINGLESTOREDB_INFERENCE_API_HOSTING_PLATFORM') |
| 70 | + if base_url is None or hosting_platform is None: |
| 71 | + inference_api_manager = ( |
| 72 | + manage_workspaces().organizations.current.inference_apis |
| 73 | + ) |
| 74 | + info = inference_api_manager.get(model_name=model_name) |
| 75 | + else: |
| 76 | + info = InferenceAPIInfo( |
| 77 | + service_id='', |
| 78 | + model_name=model_name, |
| 79 | + name='', |
| 80 | + connection_url=base_url, |
| 81 | + project_id='', |
| 82 | + hosting_platform=hosting_platform, |
| 83 | + ) |
| 84 | + if base_url is not None: |
| 85 | + info.connection_url = base_url |
| 86 | + if hosting_platform is not None: |
| 87 | + info.hosting_platform = hosting_platform |
| 88 | + |
| 89 | + # Extract timeouts from http_client if provided |
| 90 | + t = http_client.timeout if http_client is not None else None |
| 91 | + connect_timeout = None |
| 92 | + read_timeout = None |
| 93 | + if t is not None: |
| 94 | + if isinstance(t, httpx.Timeout): |
| 95 | + if t.connect is not None: |
| 96 | + connect_timeout = float(t.connect) |
| 97 | + if t.read is not None: |
| 98 | + read_timeout = float(t.read) |
| 99 | + if connect_timeout is None and read_timeout is not None: |
| 100 | + connect_timeout = read_timeout |
| 101 | + if read_timeout is None and connect_timeout is not None: |
| 102 | + read_timeout = connect_timeout |
| 103 | + elif isinstance(t, (int, float)): |
| 104 | + connect_timeout = float(t) |
| 105 | + read_timeout = float(t) |
| 106 | + if timeout is not None: |
| 107 | + connect_timeout = timeout |
| 108 | + read_timeout = timeout |
| 109 | + t = httpx.Timeout(timeout) |
| 110 | + |
| 111 | + if info.hosting_platform == 'Amazon': |
| 112 | + # Instantiate Bedrock client |
| 113 | + cfg_kwargs = { |
| 114 | + 'signature_version': UNSIGNED, |
| 115 | + 'retries': {'max_attempts': 1, 'mode': 'standard'}, |
| 116 | + } |
| 117 | + if read_timeout is not None: |
| 118 | + cfg_kwargs['read_timeout'] = read_timeout |
| 119 | + if connect_timeout is not None: |
| 120 | + cfg_kwargs['connect_timeout'] = connect_timeout |
| 121 | + |
| 122 | + cfg = Config(**cfg_kwargs) |
| 123 | + client = boto3.client( |
| 124 | + 'bedrock-runtime', |
| 125 | + endpoint_url=info.connection_url, |
| 126 | + region_name='us-east-1', |
| 127 | + aws_access_key_id='placeholder', |
| 128 | + aws_secret_access_key='placeholder', |
| 129 | + config=cfg, |
| 130 | + ) |
| 131 | + |
| 132 | + def _inject_headers(request: Any, **_ignored: Any) -> None: |
| 133 | + """Inject dynamic auth/OBO headers prior to Bedrock sending.""" |
| 134 | + if api_key_getter_fn is not None: |
| 135 | + token_val = api_key_getter_fn() |
| 136 | + if token_val: |
| 137 | + request.headers['Authorization'] = f'Bearer {token_val}' |
| 138 | + if obo_token_getter_fn is not None: |
| 139 | + obo_val = obo_token_getter_fn() |
| 140 | + if obo_val: |
| 141 | + request.headers['X-S2-OBO'] = obo_val |
| 142 | + request.headers.pop('X-Amz-Date', None) |
| 143 | + request.headers.pop('X-Amz-Security-Token', None) |
| 144 | + |
| 145 | + emitter = client._endpoint._event_emitter |
| 146 | + emitter.register_first( |
| 147 | + 'before-send.bedrock-runtime.Converse', |
| 148 | + _inject_headers, |
| 149 | + ) |
| 150 | + emitter.register_first( |
| 151 | + 'before-send.bedrock-runtime.ConverseStream', |
| 152 | + _inject_headers, |
| 153 | + ) |
| 154 | + emitter.register_first( |
| 155 | + 'before-send.bedrock-runtime.InvokeModel', |
| 156 | + _inject_headers, |
| 157 | + ) |
| 158 | + emitter.register_first( |
| 159 | + 'before-send.bedrock-runtime.InvokeModelWithResponseStream', |
| 160 | + _inject_headers, |
| 161 | + ) |
| 162 | + |
| 163 | + return ChatBedrockConverse( |
| 164 | + model_id=model_name, |
| 165 | + endpoint_url=info.connection_url, |
| 166 | + region_name='us-east-1', |
| 167 | + aws_access_key_id='placeholder', |
| 168 | + aws_secret_access_key='placeholder', |
| 169 | + disable_streaming=not streaming, |
| 170 | + client=client, |
| 171 | + **kwargs, |
| 172 | + ) |
| 173 | + |
| 174 | + def inject_auth_headers(request: httpx.Request) -> None: |
| 175 | + """Inject dynamic auth/OBO headers before request is sent.""" |
| 176 | + if api_key_getter_fn is not None: |
| 177 | + token_val = api_key_getter_fn() |
| 178 | + if token_val: |
| 179 | + request.headers['Authorization'] = f'Bearer {token_val}' |
| 180 | + if obo_token_getter_fn is not None: |
| 181 | + obo_val = obo_token_getter_fn() |
| 182 | + if obo_val: |
| 183 | + request.headers['X-S2-OBO'] = obo_val |
| 184 | + |
| 185 | + if t is not None: |
| 186 | + http_client = httpx.Client( |
| 187 | + timeout=t, |
| 188 | + event_hooks={'request': [inject_auth_headers]}, |
| 189 | + ) |
| 190 | + else: |
| 191 | + http_client = httpx.Client( |
| 192 | + timeout=httpx.Timeout(timeout=600, connect=5.0), # default OpenAI timeout |
| 193 | + event_hooks={'request': [inject_auth_headers]}, |
| 194 | + ) |
| 195 | + |
| 196 | + # OpenAI / Azure OpenAI path |
| 197 | + openai_kwargs = dict( |
| 198 | + base_url=info.connection_url, |
| 199 | + api_key='placeholder', |
| 200 | + model=model_name, |
| 201 | + streaming=streaming, |
| 202 | + http_client=http_client, |
| 203 | + ) |
| 204 | + return ChatOpenAI( |
| 205 | + **openai_kwargs, |
| 206 | + **kwargs, |
| 207 | + ) |
0 commit comments