From 58f0d2687adaa4c921080378b84b9aecef289bb7 Mon Sep 17 00:00:00 2001
From: Anmol Jaiswal <69112823+anmolg1997@users.noreply.github.com>
Date: Wed, 29 Apr 2026 10:21:47 +0530
Subject: [PATCH] core[patch]: preserve mime_type and guard chat/completions
 for openai file blocks

The base64 file branch in `_convert_openai_format_to_data_block` was
hard-coding `mime_type="application/pdf"`, while the image branch right
above used `parsed["mime_type"]`. So a non-PDF data URI (e.g. CSV)
passed via the OpenAI Chat Completions file block shape got silently
relabeled as PDF in the v1 content block, which is wrong for non-OpenAI
chat models that consume v1 blocks via shared `_normalize_messages`.

Changes:

1. Use `parsed["mime_type"]` in the base64 file branch, matching the
   image branch right above it.
2. In `convert_to_openai_data_block(api="chat/completions")`, raise a
   clear `ValueError` when MIME is not `application/pdf`, pointing the
   caller to the Responses API. This keeps Chat Completions semantics
   intact and fails fast with a friendlier error than OpenAI's raw 400.
3. Regression tests for both behaviors.

Fixes #36939
---
 .../messages/block_translators/openai.py      | 11 ++++-
 .../messages/block_translators/test_openai.py | 46 +++++++++++++++++++
 2 files changed, 56 insertions(+), 1 deletion(-)

diff --git a/libs/core/langchain_core/messages/block_translators/openai.py b/libs/core/langchain_core/messages/block_translators/openai.py
index 627459254c5ea..0efcbbfd096fc 100644
--- a/libs/core/langchain_core/messages/block_translators/openai.py
+++ b/libs/core/langchain_core/messages/block_translators/openai.py
@@ -93,6 +93,15 @@ def convert_to_openai_data_block(
         if block.get("source_type") == "base64" or "base64" in block:
             # Handle v0 format (Base64CB): {"source_type": "base64", "data": "...", ...}
             # Handle v1 format (IDCB): {"base64": "...", ...}
+            if (
+                api == "chat/completions"
+                and block.get("mime_type") != "application/pdf"
+            ):
+                error_msg = (
+                    "OpenAI Chat Completions only supports application/pdf for "
+                    "file data blocks. Use Responses API for other file MIME types."
+                )
+                raise ValueError(error_msg)
             base64_data = block["data"] if "source_type" in block else block["base64"]
             file = {"file_data": f"data:{block['mime_type']};base64,{base64_data}"}
             if filename := block.get("filename"):
@@ -545,7 +554,7 @@ def _extract_extras(block_dict: dict, known_keys: set[str]) -> dict[str, Any]:
         filename = block["file"].get("filename")
         return types.create_file_block(
             base64=parsed["data"],
-            mime_type="application/pdf",
+            mime_type=parsed["mime_type"],
             filename=filename,
             **all_extras,
         )
diff --git a/libs/core/tests/unit_tests/messages/block_translators/test_openai.py b/libs/core/tests/unit_tests/messages/block_translators/test_openai.py
index b39b768993cda..f1ab7e79d2ed3 100644
--- a/libs/core/tests/unit_tests/messages/block_translators/test_openai.py
+++ b/libs/core/tests/unit_tests/messages/block_translators/test_openai.py
@@ -490,6 +490,16 @@ def test_convert_to_openai_data_block() -> None:
     result = convert_to_openai_data_block(block)
     assert result == expected
 
+    # File / base64 with non-PDF MIME (chat/completions)
+    block = {
+        "type": "file",
+        "base64": "<base64 string>",
+        "mime_type": "text/csv",
+        "filename": "test.csv",
+    }
+    with pytest.raises(ValueError, match="only supports application/pdf"):
+        convert_to_openai_data_block(block)
+
     # Image / base64
     block = {
         "type": "image",
@@ -603,3 +613,39 @@ def test_convert_to_openai_data_block() -> None:
     expected = {"type": "input_file", "file_id": "file-abc123"}
     result = convert_to_openai_data_block(block, api="responses")
     assert result == expected
+
+
+def test_convert_to_v1_file_block_preserves_non_pdf_mime_type() -> None:
+    message = HumanMessage(
+        content=[
+            {
+                "type": "file",
+                "file": {
+                    "filename": "sheet.csv",
+                    "file_data": "data:text/csv;base64,aGVsbG8=",
+                },
+            },
+            {
+                "type": "file",
+                "file": {
+                    "file_data": "data:text/plain;base64,aGVsbG8=",
+                },
+            },
+        ]
+    )
+
+    expected: list[types.ContentBlock] = [
+        {
+            "type": "file",
+            "base64": "aGVsbG8=",
+            "mime_type": "text/csv",
+            "extras": {"filename": "sheet.csv"},
+        },
+        {
+            "type": "file",
+            "base64": "aGVsbG8=",
+            "mime_type": "text/plain",
+        },
+    ]
+
+    assert _content_blocks_equal_ignore_id(message.content_blocks, expected)