fix(anthropic): handle empty AIMessage (#33390)

Author: ccurme

libs/partners/anthropic/langchain_anthropic/chat_models.py

@@ -549,6 +549,10 @@ def _format_messages(
                 _lc_tool_calls_to_anthropic_tool_use_blocks(missing_tool_calls),
             )
 
+        if not content and role == "assistant" and _i < len(merged_messages) - 1:
+            # anthropic.BadRequestError: Error code: 400: all messages must have
+            # non-empty content except for the optional final assistant message
+            continue
         formatted_messages.append({"role": role, "content": content})
     return system, formatted_messages
 

libs/partners/anthropic/tests/integration_tests/test_chat_models.py

@@ -278,6 +278,53 @@ def test_system_invoke() -> None:
     assert isinstance(result.content, str)
 
 
+def test_handle_empty_aimessage() -> None:
+    # Anthropic can generate empty AIMessages, which are not valid unless in the last
+    # message in a sequence.
+    llm = ChatAnthropic(model=MODEL_NAME)
+    messages = [
+        HumanMessage("Hello"),
+        AIMessage([]),
+        HumanMessage("My name is Bob."),
+    ]
+    _ = llm.invoke(messages)
+
+    # Test tool call sequence
+    llm_with_tools = llm.bind_tools(
+        [
+            {
+                "name": "get_weather",
+                "description": "Get weather report for a city",
+                "input_schema": {
+                    "type": "object",
+                    "properties": {"location": {"type": "string"}},
+                },
+            },
+        ],
+    )
+    _ = llm_with_tools.invoke(
+        [
+            HumanMessage("What's the weather in Boston?"),
+            AIMessage(
+                content=[],
+                tool_calls=[
+                    {
+                        "name": "get_weather",
+                        "args": {"location": "Boston"},
+                        "id": "toolu_01V6d6W32QGGSmQm4BT98EKk",
+                        "type": "tool_call",
+                    },
+                ],
+            ),
+            ToolMessage(
+                content="It's sunny.", tool_call_id="toolu_01V6d6W32QGGSmQm4BT98EKk"
+            ),
+            AIMessage([]),
+            HumanMessage("Thanks!"),
+        ]
+    )
+
+
 def test_anthropic_call() -> None:
     """Test valid call to anthropic."""
     chat = ChatAnthropic(model=MODEL_NAME)  # type: ignore[call-arg]

libs/partners/anthropic/tests/unit_tests/test_chat_models.py

@@ -642,6 +642,38 @@ def test__format_messages_with_tool_calls() -> None:
     actual = _format_messages(messages)
     assert expected == actual
 
+    # Check handling of empty AIMessage
+    empty_contents: list[str | list[str | dict]] = ["", []]
+    for empty_content in empty_contents:
+        ## Permit message in final position
+        _, anthropic_messages = _format_messages([human, AIMessage(empty_content)])
+        expected_messages = [
+            {"role": "user", "content": "foo"},
+            {"role": "assistant", "content": empty_content},
+        ]
+        assert expected_messages == anthropic_messages
+
+        ## Remove message otherwise
+        _, anthropic_messages = _format_messages(
+            [human, AIMessage(empty_content), human]
+        )
+        expected_messages = [
+            {"role": "user", "content": "foo"},
+            {"role": "user", "content": "foo"},
+        ]
+        assert expected_messages == anthropic_messages
+
+        actual = _format_messages(
+            [system, human, ai, tool, AIMessage(empty_content), human]
+        )
+        assert actual[0] == "fuzz"
+        assert [message["role"] for message in actual[1]] == [
+            "user",
+            "assistant",
+            "user",
+            "user",
+        ]
+
 
 def test__format_tool_use_block() -> None:
     # Test we correctly format tool_use blocks when there is no corresponding tool_call.