[Bugfix] Fix Anthropic API base64 image handling in Messages endpoint (#35557)

Signed-off-by: Martin Vit <martin@voipmonitor.org>

tests/entrypoints/openai/test_anthropic_messages_conversion.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Unit tests for Anthropic-to-OpenAI request conversion.
+
+Tests the image source handling and tool_result content parsing in
+AnthropicServingMessages._convert_anthropic_to_openai_request().
+"""
+
+from vllm.entrypoints.anthropic.protocol import (
+    AnthropicMessagesRequest,
+)
+from vllm.entrypoints.anthropic.serving import AnthropicServingMessages
+
+_convert = AnthropicServingMessages._convert_anthropic_to_openai_request
+_img_url = AnthropicServingMessages._convert_image_source_to_url
+
+
+def _make_request(
+    messages: list[dict],
+    **kwargs,
+) -> AnthropicMessagesRequest:
+    return AnthropicMessagesRequest(
+        model="test-model",
+        max_tokens=128,
+        messages=messages,
+        **kwargs,
+    )
+
+
+# ======================================================================
+# _convert_image_source_to_url
+# ======================================================================
+
+
+class TestConvertImageSourceToUrl:
+    def test_base64_source(self):
+        source = {
+            "type": "base64",
+            "media_type": "image/jpeg",
+            "data": "iVBORw0KGgo=",
+        }
+        assert _img_url(source) == "data:image/jpeg;base64,iVBORw0KGgo="
+
+    def test_base64_png(self):
+        source = {
+            "type": "base64",
+            "media_type": "image/png",
+            "data": "AAAA",
+        }
+        assert _img_url(source) == "data:image/png;base64,AAAA"
+
+    def test_url_source(self):
+        source = {
+            "type": "url",
+            "url": "https://example.com/image.jpg",
+        }
+        assert _img_url(source) == "https://example.com/image.jpg"
+
+    def test_missing_type_defaults_to_base64(self):
+        """When 'type' is absent, treat as base64."""
+        source = {
+            "media_type": "image/webp",
+            "data": "UklGR",
+        }
+        assert _img_url(source) == "data:image/webp;base64,UklGR"
+
+    def test_missing_media_type_defaults_to_jpeg(self):
+        source = {"type": "base64", "data": "abc123"}
+        assert _img_url(source) == "data:image/jpeg;base64,abc123"
+
+    def test_url_source_missing_url_returns_empty(self):
+        source = {"type": "url"}
+        assert _img_url(source) == ""
+
+    def test_empty_source_returns_data_uri_shell(self):
+        source: dict = {}
+        assert _img_url(source) == "data:image/jpeg;base64,"
+
+
+# ======================================================================
+# Image blocks inside user messages
+# ======================================================================
+
+
+class TestImageContentBlocks:
+    def test_base64_image_in_user_message(self):
+        request = _make_request(
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": "Describe this image"},
+                        {
+                            "type": "image",
+                            "source": {
+                                "type": "base64",
+                                "media_type": "image/jpeg",
+                                "data": "iVBORw0KGgo=",
+                            },
+                        },
+                    ],
+                }
+            ]
+        )
+
+        result = _convert(request)
+        user_msg = result.messages[0]
+        assert user_msg["role"] == "user"
+
+        parts = user_msg["content"]
+        assert len(parts) == 2
+        assert parts[0] == {"type": "text", "text": "Describe this image"}
+        assert parts[1] == {
+            "type": "image_url",
+            "image_url": {"url": "data:image/jpeg;base64,iVBORw0KGgo="},
+        }
+
+    def test_url_image_in_user_message(self):
+        request = _make_request(
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": "What is this?"},
+                        {
+                            "type": "image",
+                            "source": {
+                                "type": "url",
+                                "url": "https://example.com/cat.png",
+                            },
+                        },
+                    ],
+                }
+            ]
+        )
+
+        result = _convert(request)
+        parts = result.messages[0]["content"]
+        assert parts[1] == {
+            "type": "image_url",
+            "image_url": {"url": "https://example.com/cat.png"},
+        }
+
+
+# ======================================================================
+# tool_result content handling
+# ======================================================================
+
+
+class TestToolResultContent:
+    def _make_tool_result_request(
+        self, tool_result_content
+    ) -> AnthropicMessagesRequest:
+        """Build a request with assistant tool_use followed by user
+        tool_result."""
+        return _make_request(
+            [
+                {
+                    "role": "assistant",
+                    "content": [
+                        {
+                            "type": "tool_use",
+                            "id": "call_001",
+                            "name": "read_file",
+                            "input": {"path": "/tmp/img.png"},
+                        }
+                    ],
+                },
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "tool_result",
+                            "tool_use_id": "call_001",
+                            "content": tool_result_content,
+                        }
+                    ],
+                },
+            ]
+        )
+
+    def test_tool_result_string_content(self):
+        request = self._make_tool_result_request("file contents here")
+        result = _convert(request)
+
+        tool_msg = [m for m in result.messages if m["role"] == "tool"]
+        assert len(tool_msg) == 1
+        assert tool_msg[0]["content"] == "file contents here"
+        assert tool_msg[0]["tool_call_id"] == "call_001"
+
+    def test_tool_result_text_blocks(self):
+        request = self._make_tool_result_request(
+            [
+                {"type": "text", "text": "line 1"},
+                {"type": "text", "text": "line 2"},
+            ]
+        )
+        result = _convert(request)
+
+        tool_msg = [m for m in result.messages if m["role"] == "tool"]
+        assert len(tool_msg) == 1
+        assert tool_msg[0]["content"] == "line 1\nline 2"
+
+    def test_tool_result_with_image(self):
+        """Image in tool_result should produce a follow-up user message."""
+        request = self._make_tool_result_request(
+            [
+                {
+                    "type": "image",
+                    "source": {
+                        "type": "base64",
+                        "media_type": "image/png",
+                        "data": "AAAA",
+                    },
+                }
+            ]
+        )
+        result = _convert(request)
+
+        tool_msg = [m for m in result.messages if m["role"] == "tool"]
+        assert len(tool_msg) == 1
+        assert tool_msg[0]["content"] == ""
+
+        # The image should be injected as a follow-up user message
+        follow_up = [
+            m
+            for m in result.messages
+            if m["role"] == "user" and isinstance(m.get("content"), list)
+        ]
+        assert len(follow_up) == 1
+        img_parts = follow_up[0]["content"]
+        assert len(img_parts) == 1
+        assert img_parts[0] == {
+            "type": "image_url",
+            "image_url": {"url": "data:image/png;base64,AAAA"},
+        }
+
+    def test_tool_result_with_text_and_image(self):
+        """Mixed text+image tool_result: text in tool msg, image in user
+        msg."""
+        request = self._make_tool_result_request(
+            [
+                {"type": "text", "text": "Here is the screenshot"},
+                {
+                    "type": "image",
+                    "source": {
+                        "type": "base64",
+                        "media_type": "image/jpeg",
+                        "data": "QUFB",
+                    },
+                },
+            ]
+        )
+        result = _convert(request)
+
+        tool_msg = [m for m in result.messages if m["role"] == "tool"]
+        assert len(tool_msg) == 1
+        assert tool_msg[0]["content"] == "Here is the screenshot"
+
+        follow_up = [
+            m
+            for m in result.messages
+            if m["role"] == "user" and isinstance(m.get("content"), list)
+        ]
+        assert len(follow_up) == 1
+        assert follow_up[0]["content"][0]["image_url"]["url"] == (
+            "data:image/jpeg;base64,QUFB"
+        )
+
+    def test_tool_result_with_multiple_images(self):
+        request = self._make_tool_result_request(
+            [
+                {
+                    "type": "image",
+                    "source": {
+                        "type": "base64",
+                        "media_type": "image/png",
+                        "data": "IMG1",
+                    },
+                },
+                {
+                    "type": "image",
+                    "source": {
+                        "type": "url",
+                        "url": "https://example.com/img2.jpg",
+                    },
+                },
+            ]
+        )
+        result = _convert(request)
+
+        follow_up = [
+            m
+            for m in result.messages
+            if m["role"] == "user" and isinstance(m.get("content"), list)
+        ]
+        assert len(follow_up) == 1
+        urls = [p["image_url"]["url"] for p in follow_up[0]["content"]]
+        assert urls == [
+            "data:image/png;base64,IMG1",
+            "https://example.com/img2.jpg",
+        ]
+
+    def test_tool_result_none_content(self):
+        request = self._make_tool_result_request(None)
+        result = _convert(request)
+
+        tool_msg = [m for m in result.messages if m["role"] == "tool"]
+        assert len(tool_msg) == 1
+        assert tool_msg[0]["content"] == ""
+
+    def test_tool_result_no_follow_up_when_no_images(self):
+        """Ensure no extra user message is added when there are no images."""
+        request = self._make_tool_result_request(
+            [
+                {"type": "text", "text": "just text"},
+            ]
+        )
+        result = _convert(request)
+
+        user_follow_ups = [
+            m
+            for m in result.messages
+            if m["role"] == "user" and isinstance(m.get("content"), list)
+        ]
+        assert len(user_follow_ups) == 0

vllm/entrypoints/anthropic/serving.py

@@ -86,8 +86,30 @@ class AnthropicServingMessages(OpenAIServingChat):
             "tool_calls": "tool_use",
         }
 
+    @staticmethod
+    def _convert_image_source_to_url(source: dict[str, Any]) -> str:
+        """Convert an Anthropic image source to an OpenAI-compatible URL.
+
+        Anthropic supports two image source types:
+        - base64: {"type": "base64", "media_type": "image/jpeg", "data": "..."}
+        - url: {"type": "url", "url": "https://..."}
+
+        For base64 sources, this constructs a proper data URI that
+        downstream processors (e.g. vLLM's media connector) can handle.
+        """
+        source_type = source.get("type")
+        if source_type == "url":
+            return source.get("url", "")
+        # Default to base64 processing if type is "base64"
+        # or missing, ensuring a proper data URI is always
+        # constructed for non-URL sources.
+        media_type = source.get("media_type", "image/jpeg")
+        data = source.get("data", "")
+        return f"data:{media_type};base64,{data}"
+
+    @classmethod
     def _convert_anthropic_to_openai_request(
-        self, anthropic_request: AnthropicMessagesRequest
+        cls, anthropic_request: AnthropicMessagesRequest
     ) -> ChatCompletionRequest:
         """Convert Anthropic message format to OpenAI format"""
         openai_messages = []
@@ -119,10 +141,11 @@ class AnthropicServingMessages(OpenAIServingChat):
                     if block.type == "text" and block.text:
                         content_parts.append({"type": "text", "text": block.text})
                     elif block.type == "image" and block.source:
+                        image_url = cls._convert_image_source_to_url(block.source)
                         content_parts.append(
                             {
                                 "type": "image_url",
-                                "image_url": {"url": block.source.get("data", "")},
+                                "image_url": {"url": image_url},
                             }
                         )
                     elif block.type == "thinking" and block.thinking is not None:
@@ -140,13 +163,48 @@ class AnthropicServingMessages(OpenAIServingChat):
                         tool_calls.append(tool_call)
                     elif block.type == "tool_result":
                         if msg.role == "user":
+                            # Parse tool_result content which can be
+                            # a string or a list of content blocks
+                            # (text, image, etc.)
+                            tool_text = ""
+                            tool_image_urls: list[str] = []
+                            if isinstance(block.content, str):
+                                tool_text = block.content
+                            elif isinstance(block.content, list):
+                                text_parts: list[str] = []
+                                for item in block.content:
+                                    if not isinstance(item, dict):
+                                        continue
+                                    item_type = item.get("type")
+                                    if item_type == "text":
+                                        text_parts.append(item.get("text", ""))
+                                    elif item_type == "image":
+                                        source = item.get("source", {})
+                                        url = cls._convert_image_source_to_url(source)
+                                        if url:
+                                            tool_image_urls.append(url)
+                                tool_text = "\n".join(text_parts)
                             openai_messages.append(
                                 {
                                     "role": "tool",
                                     "tool_call_id": block.tool_use_id or "",
-                                    "content": str(block.content)
-                                    if block.content
-                                    else "",
+                                    "content": tool_text or "",
+                                }
+                            )
+                            # OpenAI tool messages only support string
+                            # content, so inject images from tool
+                            # results as a follow-up user message
+                            if tool_image_urls:
+                                openai_messages.append(
+                                    {
+                                        "role": "user",
+                                        "content": [  # type: ignore[dict-item]
+                                            {
+                                                "type": "image_url",
+                                                "image_url": {"url": img},
+                                            }
+                                            for img in tool_image_urls
+                                        ],
                                     }
                                 )
                         else: