feat: add Vertex AI support for all providers (#135)

* feat: add Vertex AI support for all providers

Route requests through Vertex AI when GoogleADC auth is provided.
Supports Google, Anthropic, Mistral, and DeepSeek providers across
text, images, and videos modalities. Includes Veo polling fix
(fetchPredictOperation), error handler hardening, Gemini image
role fix, and DeepSeek usage parser fix.

WIP: Veo Vertex inline video (bytesBase64Encoded) parsing not yet
handled — needs base64 decoding or storageUri in request.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(ci): install gcp extra so Vertex routing tests can import google-auth

The google-auth package is optional under [gcp], but unit tests in
test_vertex_routing.py and Vertex integration tests need it importable.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(veo): handle Vertex inline bytesBase64Encoded and videoGcsUri key mismatch

Vertex Veo responses use videoGcsUri (not uri/gcsUri) and can return
inline base64 instead of a GCS URL. Normalize the key and decode
inline responses directly into VideoArtifact.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(embeddings): adapt request/response format for Vertex :predict endpoint

Vertex embeddings uses :predict with instances format, not :embedContent.
Build correct request body in _init_request when auth is GoogleADC, and
parse predictions response format in _parse_content. Add integration test.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* chore: fix trailing newlines in workflow files

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* refactor(vertex): move embeddings auth to provider mixin, rename VertexEndpoint, update templates

- Move isinstance(self.auth, GoogleADC) check from modality _init_request() to
  provider mixin _make_request() for embeddings, keeping auth logic in provider layer
- Fix misplaced class docstring in GoogleEmbeddingsClient mixin
- Rename VertexEndpoint to VertexGenerateContentEndpoint for consistency with
  VertexImagenEndpoint, VertexEmbeddingsEndpoint, etc.
- Add Vertex AI routing patterns (commented) to provider templates

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* refactor(vertex): centralize URL building in GoogleADC.build_url()

Move duplicated project_id validation, base URL resolution, and endpoint
formatting from 7 provider _build_url() methods into GoogleADC.build_url().
Also remove manual base64.b64decode from video client (Artifact validator
handles it).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(images): let Artifact validator handle base64 decoding in Gemini images

Same pattern as the video client fix - pass base64 string directly to
ImageArtifact(data=...) instead of manual base64.b64decode().

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(images): let Artifact validator handle base64 decoding in Imagen

Same pattern as Gemini images and Veo video fixes - pass base64 string
directly to ImageArtifact(data=...) instead of manual base64.b64decode().

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: Kamilbenkirane

.github/workflows/ci.yml

@@ -85,6 +85,7 @@ jobs:
       - uses: ./.github/actions/setup-python-uv
         with:
           python-version: ${{ matrix.python-version }}
+      - run: uv sync --extra gcp
       - run: uv run pytest tests/unit_tests -v --cov=celeste --cov-report=term-missing --cov-report=xml --cov-report=html --cov-fail-under=80
       - uses: codecov/codecov-action@v4
         if: matrix.os == 'ubuntu-latest' && matrix.python-version == '3.12'

.github/workflows/claude-code-review.yml

@@ -41,4 +41,3 @@ jobs:
           prompt: '/code-review:code-review ${{ github.repository }}/pull/${{ github.event.pull_request.number }}'
           # See https://github.com/anthropics/claude-code-action/blob/main/docs/usage.md
           # or https://code.claude.com/docs/en/cli-reference for available options
-

.github/workflows/claude.yml

@@ -47,4 +47,3 @@ jobs:
           # See https://github.com/anthropics/claude-code-action/blob/main/docs/usage.md
           # or https://code.claude.com/docs/en/cli-reference for available options
           # claude_args: '--allowed-tools Bash(gh pr:*)'
-

.github/workflows/publish.yml

@@ -58,6 +58,7 @@ jobs:
           workload_identity_provider: ${{ secrets.GCP_WORKLOAD_IDENTITY_PROVIDER }}
           service_account: ${{ secrets.GCP_SERVICE_ACCOUNT }}
       - uses: ./.github/actions/setup-python-uv
+      - run: uv sync --extra gcp
       - name: Run ${{ matrix.package }} integration tests
         env:
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

src/celeste/client.py

@@ -296,9 +296,8 @@ def _handle_error_response(self, response: httpx.Response) -> None:
         """Handle error responses from provider APIs."""
         if not response.is_success:
             try:
-                error_data = response.json()
-                error_msg = error_data.get("error", {}).get("message", response.text)
-            except JSONDecodeError:
+                error_msg = response.json()["error"]["message"]
+            except (JSONDecodeError, KeyError, TypeError, IndexError):
                 error_msg = response.text or f"HTTP {response.status_code}"
 
             raise httpx.HTTPStatusError(

src/celeste/modalities/images/providers/google/gemini.py

@@ -82,7 +82,7 @@ def _init_request(self, inputs: ImageInput) -> dict[str, Any]:
         parts.append({"text": inputs.prompt})
 
         return {
-            "contents": [{"parts": parts}],
+            "contents": [{"role": "user", "parts": parts}],
             "generationConfig": {
                 "responseModalities": ["TEXT", "IMAGE"],
                 "imageConfig": {},
@@ -113,8 +113,7 @@ def _parse_content(
                 if not base64_data:
                     continue
                 mime_type = ImageMimeType(inline_data.get("mimeType", "image/png"))
-                image_bytes = base64.b64decode(base64_data)
-                artifacts.append(ImageArtifact(data=image_bytes, mime_type=mime_type))
+                artifacts.append(ImageArtifact(data=base64_data, mime_type=mime_type))
 
         if not artifacts:
             return ImageArtifact()

src/celeste/modalities/images/providers/google/imagen.py

@@ -1,6 +1,5 @@
 """Imagen client for Google images modality."""
 
-import base64
 from typing import Any, Unpack
 
 from celeste.artifacts import ImageArtifact
@@ -61,8 +60,7 @@ def _parse_content(
             if not base64_data:
                 continue
             mime_type = ImageMimeType(prediction.get("mimeType", "image/png"))
-            image_bytes = base64.b64decode(base64_data)
-            images.append(ImageArtifact(data=image_bytes, mime_type=mime_type))
+            images.append(ImageArtifact(data=base64_data, mime_type=mime_type))
 
         num_images_requested = parameters.get("num_images")
         if num_images_requested == 1:

src/celeste/modalities/videos/providers/google/client.py

@@ -3,6 +3,7 @@
 from typing import Any, Unpack
 
 from celeste.artifacts import VideoArtifact
+from celeste.mime_types import VideoMimeType
 from celeste.parameters import ParameterMapper
 from celeste.providers.google.veo import config
 from celeste.providers.google.veo.client import GoogleVeoClient as GoogleVeoMixin
@@ -51,6 +52,12 @@ def _parse_content(
     ) -> VideoArtifact:
         """Parse content from response."""
         video_data = super()._parse_content(response_data)
+        # Handle inline base64 response (Vertex can return bytesBase64Encoded)
+        if "bytesBase64Encoded" in video_data:
+            mime_type = video_data.get("mimeType", VideoMimeType.MP4)
+            return VideoArtifact(
+                data=video_data["bytesBase64Encoded"], mime_type=mime_type
+            )
         return VideoArtifact(url=video_data.get("uri"))
 
     def _parse_finish_reason(self, response_data: dict[str, Any]) -> VideoFinishReason:
@@ -62,13 +69,16 @@ async def download_content(self, artifact: VideoArtifact) -> VideoArtifact:
         """Download video content from GCS URL.
 
         Args:
-            artifact: VideoArtifact with URL to download.
+            artifact: VideoArtifact with URL or inline data to download.
 
         Returns:
             VideoArtifact with downloaded bytes data.
         """
+        if artifact.data is not None:
+            return artifact
+
         if artifact.url is None:
-            msg = "Artifact has no URL to download"
+            msg = "Artifact has no URL or data to download"
             raise ValueError(msg)
 
         video_bytes = await super().download_content(artifact.url)

src/celeste/providers/anthropic/messages/client.py

@@ -7,6 +7,7 @@
 from celeste.core import UsageField
 from celeste.io import FinishReason
 from celeste.mime_types import ApplicationMimeType
+from celeste.providers.google.auth import GoogleADC
 
 from . import config
 
@@ -22,6 +23,10 @@ class AnthropicMessagesClient(APIMixin):
     - _parse_finish_reason() - Extract finish reason from response
     - _build_metadata() - Filter content fields
 
+    Auth-based endpoint selection:
+    - GoogleADC auth -> Vertex AI endpoints (Claude on Google Cloud)
+    - API key auth -> Anthropic API endpoints
+
     Usage:
         class AnthropicTextGenerationClient(AnthropicMessagesClient, TextGenerationClient):
             def _parse_content(self, response_data, **parameters):
@@ -32,6 +37,23 @@ def _parse_content(self, response_data, **parameters):
                 return ""
     """
 
+    def _get_vertex_endpoint(
+        self, anthropic_endpoint: str, streaming: bool = False
+    ) -> str:
+        """Map Anthropic endpoint to Vertex AI endpoint."""
+        if streaming:
+            return config.VertexAnthropicEndpoint.STREAM_MESSAGE
+        return config.VertexAnthropicEndpoint.CREATE_MESSAGE
+
+    def _build_url(self, endpoint: str, streaming: bool = False) -> str:
+        """Build full URL based on auth type."""
+        if isinstance(self.auth, GoogleADC):
+            return self.auth.build_url(
+                self._get_vertex_endpoint(endpoint, streaming=streaming),
+                model_id=self.model.id,
+            )
+        return f"{config.BASE_URL}{endpoint}"
+
     def _build_headers(self, request_body: dict[str, Any]) -> dict[str, str]:
         """Build headers with beta features extracted from request."""
         beta_features: list[str] = request_body.pop("_beta_features", [])
@@ -85,7 +107,7 @@ async def _make_request(
             endpoint = config.AnthropicMessagesEndpoint.CREATE_MESSAGE
 
         response = await self.http_client.post(
-            f"{config.BASE_URL}{endpoint}",
+            url=self._build_url(endpoint, streaming=False),
             headers=headers,
             json_body=request_body,
         )
@@ -111,7 +133,7 @@ def _make_stream_request(
             endpoint = config.AnthropicMessagesEndpoint.CREATE_MESSAGE
 
         return self.http_client.stream_post(
-            f"{config.BASE_URL}{endpoint}",
+            url=self._build_url(endpoint, streaming=True),
             headers=headers,
             json_body=request_body,
         )

src/celeste/providers/anthropic/messages/config.py

@@ -12,6 +12,13 @@ class AnthropicMessagesEndpoint(StrEnum):
     GET_MODEL = "/v1/models/{model_id}"
 
 
+class VertexAnthropicEndpoint(StrEnum):
+    """Endpoints for Anthropic on Vertex AI."""
+
+    CREATE_MESSAGE = "/v1/projects/{project_id}/locations/{location}/publishers/anthropic/models/{model_id}:rawPredict"
+    STREAM_MESSAGE = "/v1/projects/{project_id}/locations/{location}/publishers/anthropic/models/{model_id}:streamRawPredict"
+
+
 BASE_URL = "https://api.anthropic.com"
 
 # Required