fix(core): (v1) trace PDFs in v0 standard format (#33009)

libs/core/langchain_core/language_models/chat_models.py

@@ -48,7 +48,6 @@ from langchain_core.messages import (
     message_chunk_to_message,
 )
 from langchain_core.messages.block_translators.openai import (
-    convert_to_openai_data_block,
     convert_to_openai_image_block,
 )
 from langchain_core.output_parsers.openai_tools import (
@@ -149,16 +148,18 @@ def _format_for_tracing(messages: list[BaseMessage]) -> list[BaseMessage]:
                         block.get("type") == "file"
                         and is_data_content_block(block)  # v0 (image/audio/file) or v1
                         and "base64" in block
-                        # Narrows to old Base64ContentBlock or new FileContentBlock
+                        # Backward compat: convert v1 base64 blocks to v0
                     ):
                         if message_to_trace is message:
                             # Shallow copy
                             message_to_trace = message.model_copy()
                             message_to_trace.content = list(message_to_trace.content)
 
-                        message_to_trace.content[idx] = convert_to_openai_data_block(  # type: ignore[index]
-                            block
-                        )
+                        message_to_trace.content[idx] = {  # type: ignore[index]
+                            **{k: v for k, v in block.items() if k != "base64"},
+                            "data": block["base64"],
+                            "source_type": "base64",
+                        }
                     elif len(block) == 1 and "type" not in block:
                         # Tracing assumes all content blocks have a "type" key. Here
                         # we add this key if it is missing, and there's an obvious

libs/core/tests/unit_tests/language_models/chat_models/test_base.py

@@ -1,6 +1,7 @@
 """Test base chat model."""
 
 import uuid
+import warnings
 from collections.abc import AsyncIterator, Iterator
 from typing import TYPE_CHECKING, Any, Literal, Optional, Union
 
@@ -505,6 +506,45 @@ def test_trace_images_in_openai_format() -> None:
     ]
 
 
+def test_trace_pdfs() -> None:
+    # For backward compat
+    llm = ParrotFakeChatModel()
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "file",
+                    "mime_type": "application/pdf",
+                    "base64": "<base64 string>",
+                }
+            ],
+        }
+    ]
+    tracer = FakeChatModelStartTracer()
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        llm.invoke(messages, config={"callbacks": [tracer]})
+
+    assert tracer.messages == [
+        [
+            [
+                HumanMessage(
+                    content=[
+                        {
+                            "type": "file",
+                            "mime_type": "application/pdf",
+                            "source_type": "base64",
+                            "data": "<base64 string>",
+                        }
+                    ]
+                )
+            ]
+        ]
+    ]
+
+
 def test_content_block_transformation_v0_to_v1_image() -> None:
     """Test that v0 format image content blocks are transformed to v1 format."""
     # Create a message with v0 format image content

libs/partners/openai/tests/unit_tests/chat_models/test_base.py

@@ -789,6 +789,22 @@ def test_format_message_content() -> None:
     for content in contents:
         assert expected == _format_message_content([content])
 
+    # Test warn if PDF is missing a filename
+    pdf_block = {
+        "type": "file",
+        "base64": "<base64 data>",
+        "mime_type": "application/pdf",
+    }
+    expected = [
+        # N.B. this format is invalid for OpenAI
+        {
+            "type": "file",
+            "file": {"file_data": "data:application/pdf;base64,<base64 data>"},
+        }
+    ]
+    with pytest.warns(match="filename"):
+        assert expected == _format_message_content([pdf_block])
+
     contents = [
         {"type": "file", "source_type": "id", "id": "file-abc123"},
         {"type": "file", "file_id": "file-abc123"},