|
| 1 | +# Copyright (c) 2025 Oracle and/or its affiliates. |
| 2 | +# Licensed under the Universal Permissive License v 1.0 as shown at https://oss.oracle.com/licenses/upl/ |
| 3 | + |
| 4 | +from langchain_core.embeddings import Embeddings |
| 5 | +from langchain_core.language_models.llms import create_base_retry_decorator |
| 6 | +from langchain_core.utils import get_from_dict_or_env |
| 7 | +from pydantic import BaseModel, Field, model_validator |
| 8 | +import requests |
| 9 | +from typing import Any, Callable, Dict, List, Mapping, Optional |
| 10 | + |
| 11 | + |
| 12 | +DEFAULT_HEADER = { |
| 13 | + "Content-Type": "application/json", |
| 14 | +} |
| 15 | + |
| 16 | + |
| 17 | +class TokenExpiredError(Exception): |
| 18 | + pass |
| 19 | + |
| 20 | + |
| 21 | +def _create_retry_decorator(llm) -> Callable[[Any], Any]: |
| 22 | + """Creates a retry decorator.""" |
| 23 | + errors = [requests.exceptions.ConnectTimeout, TokenExpiredError] |
| 24 | + decorator = create_base_retry_decorator( |
| 25 | + error_types=errors, max_retries=llm.max_retries |
| 26 | + ) |
| 27 | + return decorator |
| 28 | + |
| 29 | + |
| 30 | +class OCIModelDeploymentEndpointEmbeddings(BaseModel, Embeddings): |
| 31 | + """Embedding model deployed on OCI Data Science Model Deployment. |
| 32 | +
|
| 33 | + Example: |
| 34 | +
|
| 35 | + .. code-block:: python |
| 36 | +
|
| 37 | + from langchain_oci.embeddings import OCIModelDeploymentEndpointEmbeddings |
| 38 | +
|
| 39 | + embeddings = OCIModelDeploymentEndpointEmbeddings( |
| 40 | + endpoint="https://modeldeployment.us-ashburn-1.oci.customer-oci.com/<md_ocid>/predict", |
| 41 | + ) |
| 42 | + """ # noqa: E501 |
| 43 | + |
| 44 | + auth: dict = Field(default_factory=dict, exclude=True) |
| 45 | + """ADS auth dictionary for OCI authentication: |
| 46 | + https://accelerated-data-science.readthedocs.io/en/latest/user_guide/cli/authentication.html. |
| 47 | + This can be generated by calling `ads.common.auth.api_keys()` |
| 48 | + or `ads.common.auth.resource_principal()`. If this is not |
| 49 | + provided then the `ads.common.default_signer()` will be used.""" |
| 50 | + |
| 51 | + endpoint: str = "" |
| 52 | + """The uri of the endpoint from the deployed Model Deployment model.""" |
| 53 | + |
| 54 | + model_kwargs: Optional[Dict] = None |
| 55 | + """Keyword arguments to pass to the model.""" |
| 56 | + |
| 57 | + endpoint_kwargs: Optional[Dict] = None |
| 58 | + """Optional attributes (except for headers) passed to the request.post |
| 59 | + function. |
| 60 | + """ |
| 61 | + |
| 62 | + max_retries: int = 1 |
| 63 | + """The maximum number of retries to make when generating.""" |
| 64 | + |
| 65 | + @model_validator(mode="before") |
| 66 | + def validate_environment( # pylint: disable=no-self-argument |
| 67 | + cls, values: Dict |
| 68 | + ) -> Dict: |
| 69 | + """Validate that python package exists in environment.""" |
| 70 | + try: |
| 71 | + import ads |
| 72 | + |
| 73 | + except ImportError as ex: |
| 74 | + raise ImportError( |
| 75 | + "Could not import ads python package. " |
| 76 | + "Please install it with `pip install oracle_ads`." |
| 77 | + ) from ex |
| 78 | + if not values.get("auth", None): |
| 79 | + values["auth"] = ads.common.auth.default_signer() |
| 80 | + values["endpoint"] = get_from_dict_or_env( |
| 81 | + values, |
| 82 | + "endpoint", |
| 83 | + "OCI_LLM_ENDPOINT", |
| 84 | + ) |
| 85 | + return values |
| 86 | + |
| 87 | + @property |
| 88 | + def _identifying_params(self) -> Mapping[str, Any]: |
| 89 | + """Get the identifying parameters.""" |
| 90 | + _model_kwargs = self.model_kwargs or {} |
| 91 | + return { |
| 92 | + **{"endpoint": self.endpoint}, |
| 93 | + **{"model_kwargs": _model_kwargs}, |
| 94 | + } |
| 95 | + |
| 96 | + def _embed_with_retry(self, **kwargs) -> Any: |
| 97 | + """Use tenacity to retry the call.""" |
| 98 | + retry_decorator = _create_retry_decorator(self) |
| 99 | + |
| 100 | + @retry_decorator |
| 101 | + def _completion_with_retry(**kwargs: Any) -> Any: |
| 102 | + try: |
| 103 | + response = requests.post(self.endpoint, **kwargs) |
| 104 | + response.raise_for_status() |
| 105 | + return response |
| 106 | + except requests.exceptions.HTTPError as http_err: |
| 107 | + if response.status_code == 401 and self._refresh_signer(): |
| 108 | + raise TokenExpiredError() from http_err |
| 109 | + else: |
| 110 | + raise ValueError( |
| 111 | + f"Server error: {str(http_err)}. Message: {response.text}" |
| 112 | + ) from http_err |
| 113 | + except Exception as e: |
| 114 | + raise ValueError(f"Error occurs by inference endpoint: {str(e)}") from e |
| 115 | + |
| 116 | + return _completion_with_retry(**kwargs) |
| 117 | + |
| 118 | + def _embedding(self, texts: List[str]) -> List[List[float]]: |
| 119 | + """Call out to OCI Data Science Model Deployment Endpoint. |
| 120 | +
|
| 121 | + Args: |
| 122 | + texts: A list of texts to embed. |
| 123 | +
|
| 124 | + Returns: |
| 125 | + A list of list of floats representing the embeddings, or None if an |
| 126 | + error occurs. |
| 127 | + """ |
| 128 | + _model_kwargs = self.model_kwargs or {} |
| 129 | + body = self._construct_request_body(texts, _model_kwargs) |
| 130 | + request_kwargs = self._construct_request_kwargs(body) |
| 131 | + response = self._embed_with_retry(**request_kwargs) |
| 132 | + return self._proceses_response(response) |
| 133 | + |
| 134 | + def _construct_request_kwargs(self, body: Any) -> dict: |
| 135 | + """Constructs the request kwargs as a dictionary.""" |
| 136 | + from ads.model.common.utils import _is_json_serializable |
| 137 | + |
| 138 | + _endpoint_kwargs = self.endpoint_kwargs or {} |
| 139 | + headers = _endpoint_kwargs.pop("headers", DEFAULT_HEADER) |
| 140 | + return ( |
| 141 | + dict( |
| 142 | + headers=headers, |
| 143 | + json=body, |
| 144 | + auth=self.auth.get("signer"), |
| 145 | + **_endpoint_kwargs, |
| 146 | + ) |
| 147 | + if _is_json_serializable(body) |
| 148 | + else dict( |
| 149 | + headers=headers, |
| 150 | + data=body, |
| 151 | + auth=self.auth.get("signer"), |
| 152 | + **_endpoint_kwargs, |
| 153 | + ) |
| 154 | + ) |
| 155 | + |
| 156 | + def _construct_request_body(self, texts: List[str], params: dict) -> Any: |
| 157 | + """Constructs the request body.""" |
| 158 | + return {"input": texts} |
| 159 | + |
| 160 | + def _proceses_response(self, response: requests.Response) -> List[List[float]]: |
| 161 | + """Extracts results from requests.Response.""" |
| 162 | + try: |
| 163 | + res_json = response.json() |
| 164 | + embeddings = res_json["data"][0]["embedding"] |
| 165 | + except Exception as e: |
| 166 | + raise ValueError( |
| 167 | + f"Error raised by inference API: {e}.\nResponse: {response.text}" |
| 168 | + ) |
| 169 | + return embeddings |
| 170 | + |
| 171 | + def embed_documents( |
| 172 | + self, |
| 173 | + texts: List[str], |
| 174 | + chunk_size: Optional[int] = None, |
| 175 | + ) -> List[List[float]]: |
| 176 | + """Compute doc embeddings using OCI Data Science Model Deployment Endpoint. |
| 177 | +
|
| 178 | + Args: |
| 179 | + texts: The list of texts to embed. |
| 180 | + chunk_size: The chunk size defines how many input texts will |
| 181 | + be grouped together as request. If None, will use the |
| 182 | + chunk size specified by the class. |
| 183 | +
|
| 184 | + Returns: |
| 185 | + List of embeddings, one for each text. |
| 186 | + """ |
| 187 | + results = [] |
| 188 | + _chunk_size = ( |
| 189 | + len(texts) if (not chunk_size or chunk_size > len(texts)) else chunk_size |
| 190 | + ) |
| 191 | + for i in range(0, len(texts), _chunk_size): |
| 192 | + response = self._embedding(texts[i : i + _chunk_size]) |
| 193 | + results.extend(response) |
| 194 | + return results |
| 195 | + |
| 196 | + def embed_query(self, text: str) -> List[float]: |
| 197 | + """Compute query embeddings using OCI Data Science Model Deployment Endpoint. |
| 198 | +
|
| 199 | + Args: |
| 200 | + text: The text to embed. |
| 201 | +
|
| 202 | + Returns: |
| 203 | + Embeddings for the text. |
| 204 | + """ |
| 205 | + return self._embedding([text])[0] |
0 commit comments