|
| 1 | +"""Anthropic Vertex AI client implementation.""" |
| 2 | + |
| 3 | +import os |
| 4 | +from collections.abc import Sequence |
| 5 | +from contextvars import ContextVar |
| 6 | +from dataclasses import replace |
| 7 | +from functools import lru_cache |
| 8 | +from typing import Literal |
| 9 | + |
| 10 | +from anthropic import NOT_GIVEN |
| 11 | +from anthropic.lib.vertex._client import AnthropicVertex, AsyncAnthropicVertex |
| 12 | + |
| 13 | +from ...content import Image |
| 14 | +from ...messages import Message, UserMessage |
| 15 | +from ..anthropic import BaseAnthropicClient |
| 16 | + |
| 17 | +ANTHROPIC_VERTEX_CLIENT_CONTEXT: ContextVar["AnthropicVertexClient | None"] = ( |
| 18 | + ContextVar("ANTHROPIC_VERTEX_CLIENT_CONTEXT", default=None) |
| 19 | +) |
| 20 | + |
| 21 | + |
| 22 | +@lru_cache(maxsize=256) |
| 23 | +def _anthropic_vertex_singleton( |
| 24 | + project_id: str | None, |
| 25 | + region: str | None, |
| 26 | +) -> "AnthropicVertexClient": |
| 27 | + """Return a cached AnthropicVertexClient instance for the given parameters.""" |
| 28 | + return AnthropicVertexClient( |
| 29 | + project_id=project_id, |
| 30 | + region=region, |
| 31 | + ) |
| 32 | + |
| 33 | + |
| 34 | +def client( |
| 35 | + *, |
| 36 | + project_id: str | None = None, |
| 37 | + region: str | None = None, |
| 38 | +) -> "AnthropicVertexClient": |
| 39 | + """Return an `AnthropicVertexClient`. |
| 40 | +
|
| 41 | + Args: |
| 42 | + project_id: GCP project ID. If None, uses GOOGLE_CLOUD_PROJECT, GCLOUD_PROJECT, |
| 43 | + CLOUD_ML_PROJECT_ID, or GCP_PROJECT_ID env vars (in that order). |
| 44 | + region: GCP region. If None, uses CLOUD_ML_REGION, GOOGLE_CLOUD_REGION, or |
| 45 | + GOOGLE_CLOUD_LOCATION env vars (in that order). |
| 46 | +
|
| 47 | + Returns: |
| 48 | + An `AnthropicVertexClient` instance. |
| 49 | +
|
| 50 | + Examples: |
| 51 | + # Use environment variables |
| 52 | + client = client() |
| 53 | +
|
| 54 | + # Use explicit parameters |
| 55 | + client = client( |
| 56 | + project_id="my-gcp-project", |
| 57 | + region="us-central1" |
| 58 | + ) |
| 59 | + """ |
| 60 | + project_id = ( |
| 61 | + project_id |
| 62 | + or os.getenv("GOOGLE_CLOUD_PROJECT") |
| 63 | + or os.getenv("GCLOUD_PROJECT") |
| 64 | + or os.getenv("CLOUD_ML_PROJECT_ID") |
| 65 | + or os.getenv("GCP_PROJECT_ID") |
| 66 | + ) |
| 67 | + region = ( |
| 68 | + region |
| 69 | + or os.getenv("CLOUD_ML_REGION") |
| 70 | + or os.getenv("GOOGLE_CLOUD_REGION") |
| 71 | + or os.getenv("GOOGLE_CLOUD_LOCATION") |
| 72 | + ) |
| 73 | + |
| 74 | + return _anthropic_vertex_singleton( |
| 75 | + project_id, |
| 76 | + region, |
| 77 | + ) |
| 78 | + |
| 79 | + |
| 80 | +def clear_cache() -> None: |
| 81 | + """Clear the client singleton cache. |
| 82 | +
|
| 83 | + This is useful for testing or when you need to force recreation |
| 84 | + of clients with updated configuration. |
| 85 | + """ |
| 86 | + _anthropic_vertex_singleton.cache_clear() |
| 87 | + |
| 88 | + |
| 89 | +def get_client() -> "AnthropicVertexClient": |
| 90 | + """Retrieve the current Anthropic Vertex client from context, or a global default. |
| 91 | +
|
| 92 | + Returns: |
| 93 | + The current Anthropic Vertex client from context if available, otherwise |
| 94 | + a global default client based on environment variables. |
| 95 | + """ |
| 96 | + ctx_client = ANTHROPIC_VERTEX_CLIENT_CONTEXT.get() |
| 97 | + return ctx_client or client() |
| 98 | + |
| 99 | + |
| 100 | +class AnthropicVertexClient( |
| 101 | + BaseAnthropicClient[AnthropicVertex, AsyncAnthropicVertex, "AnthropicVertexClient"] |
| 102 | +): |
| 103 | + """Anthropic Vertex AI client that inherits from BaseAnthropicClient. |
| 104 | +
|
| 105 | + Only overrides initialization to use Vertex-specific SDK classes and |
| 106 | + provider naming to return 'anthropic-vertex'. |
| 107 | + """ |
| 108 | + |
| 109 | + @property |
| 110 | + def _context_var(self) -> ContextVar["AnthropicVertexClient | None"]: |
| 111 | + return ANTHROPIC_VERTEX_CLIENT_CONTEXT |
| 112 | + |
| 113 | + def __init__( |
| 114 | + self, |
| 115 | + *, |
| 116 | + project_id: str | None = None, |
| 117 | + region: str | None = None, |
| 118 | + ) -> None: |
| 119 | + """Initialize the Anthropic Vertex AI client. |
| 120 | +
|
| 121 | + Args: |
| 122 | + project_id: GCP project ID. |
| 123 | + region: GCP region for Vertex AI. |
| 124 | + """ |
| 125 | + self.client = AnthropicVertex( |
| 126 | + project_id=project_id or NOT_GIVEN, |
| 127 | + region=region or NOT_GIVEN, |
| 128 | + ) |
| 129 | + self.async_client = AsyncAnthropicVertex( |
| 130 | + project_id=project_id or NOT_GIVEN, |
| 131 | + region=region or NOT_GIVEN, |
| 132 | + ) |
| 133 | + |
| 134 | + @property |
| 135 | + def provider(self) -> Literal["anthropic-vertex"]: |
| 136 | + """Return the provider name for Anthropic Vertex AI.""" |
| 137 | + return "anthropic-vertex" |
| 138 | + |
| 139 | + def _prepare_messages(self, messages: Sequence[Message]) -> Sequence[Message]: |
| 140 | + return _ensure_base64_images(messages) |
| 141 | + |
| 142 | + async def _prepare_messages_async( |
| 143 | + self, messages: Sequence[Message] |
| 144 | + ) -> Sequence[Message]: |
| 145 | + return await _ensure_base64_images_async(messages) |
| 146 | + |
| 147 | + |
| 148 | +def _ensure_base64_images(messages: Sequence[Message]) -> Sequence[Message]: |
| 149 | + """Convert URL-sourced images to base64 for Vertex AI.""" |
| 150 | + updated_messages: list[Message] = [] |
| 151 | + any_updates = False |
| 152 | + |
| 153 | + for message in messages: |
| 154 | + if isinstance(message, UserMessage): |
| 155 | + converted_content = [] |
| 156 | + content_changed = False |
| 157 | + for part in message.content: |
| 158 | + if isinstance(part, Image) and part.source.type == "url_image_source": |
| 159 | + converted_content.append(Image.download(part.source.url)) |
| 160 | + content_changed = True |
| 161 | + else: |
| 162 | + converted_content.append(part) |
| 163 | + if content_changed: |
| 164 | + message = replace(message, content=converted_content) |
| 165 | + any_updates = True |
| 166 | + updated_messages.append(message) |
| 167 | + |
| 168 | + return updated_messages if any_updates else messages |
| 169 | + |
| 170 | + |
| 171 | +async def _ensure_base64_images_async( |
| 172 | + messages: Sequence[Message], |
| 173 | +) -> Sequence[Message]: |
| 174 | + """Convert URL-sourced images to base64 for Vertex AI (async).""" |
| 175 | + updated_messages: list[Message] = [] |
| 176 | + any_updates = False |
| 177 | + |
| 178 | + for message in messages: |
| 179 | + if isinstance(message, UserMessage): |
| 180 | + converted_content = [] |
| 181 | + content_changed = False |
| 182 | + for part in message.content: |
| 183 | + if isinstance(part, Image) and part.source.type == "url_image_source": |
| 184 | + converted_content.append( |
| 185 | + await Image.download_async(part.source.url) |
| 186 | + ) |
| 187 | + content_changed = True |
| 188 | + else: |
| 189 | + converted_content.append(part) |
| 190 | + if content_changed: |
| 191 | + message = replace(message, content=converted_content) |
| 192 | + any_updates = True |
| 193 | + updated_messages.append(message) |
| 194 | + |
| 195 | + return updated_messages if any_updates else messages |
0 commit comments