Merge pull request #42267 from BerriAI/litellm_otel_v2_langfuse_ocr_output

fix(otel v2): map OCR page markdown onto the generation output

litellm/integrations/otel/model/payloads.py

@@ -427,7 +427,7 @@ class LLMCallSpanData:
         # plain ``.get`` — no repeated ``isinstance`` guards.
         raw_response: Final = payload.get("response")
         response: Final = cast(Mapping[str, object], raw_response if isinstance(raw_response, dict) else {})
-        choices_out: Final = _dicts(response.get("choices")) or _responses_choices(response)
+        choices_out: Final = _dicts(response.get("choices")) or _responses_choices(response) or _ocr_choices(response)
         # ``finish_reasons`` is metadata, not content, so derive it from
         # ``choices_out`` before gating. The raw message/choice bodies are only
         # retained when content capture is enabled (see ``capture_span_content``);
@@ -752,6 +752,22 @@ def _responses_choices(response: Mapping[str, object]) -> tuple[_Choice, ...]:
     return (choice,)
 
 
+def _ocr_choices(response: Mapping[str, object]) -> tuple[_Choice, ...]:
+    markdowns: Final = tuple(
+        text for page in _dicts(response.get("pages")) if (text := as_str(page.get("markdown"))) is not None
+    )
+    if not markdowns:
+        return ()
+    message: Final[_AssistantMessage] = {
+        "role": "assistant",
+        "content": "\n\n".join(markdowns),
+        "refusal": None,
+        "tool_calls": None,
+    }
+    choice: Final[_Choice] = {"message": message, "finish_reason": None}
+    return (choice,)
+
+
 def _responses_parts_text(parts: tuple[Mapping[str, object], ...], part_type: str, field: str) -> str | None:
     texts: Final = tuple(
         text for part in parts if part.get("type") == part_type if (text := as_str(part.get(field))) is not None

tests/test_litellm/integrations/otel/test_otel_v2_sources_of_truth.py

@@ -872,6 +872,45 @@ def test_chat_choices_win_over_a_responses_output_list():
     assert data.finish_reasons == ("stop",)
 
 
+def _ocr_payload(pages: list[object]):
+    return _sample_payload(
+        call_type="aocr",
+        custom_llm_provider="mistral",
+        model="mistral-ocr-latest",
+        messages=None,
+        response={"object": "ocr", "model": "mistral-ocr-latest", "pages": pages, "usage_info": {"pages_processed": 2}},
+    )
+
+
+def test_ocr_pages_become_one_assistant_choice_joined_in_page_order():
+    data = LLMCallSpanData.from_standard_logging_payload(
+        _ocr_payload([{"index": 0, "markdown": "# Invoice"}, {"index": 1, "markdown": "Total: 42"}]),
+        capture_content=True,
+    )
+
+    assert data.choices_out == (
+        {
+            "message": {"role": "assistant", "content": "# Invoice\n\nTotal: 42", "refusal": None, "tool_calls": None},
+            "finish_reason": None,
+        },
+    )
+    assert data.finish_reasons == ()
+
+
+def test_ocr_output_follows_the_content_capture_gate():
+    data = LLMCallSpanData.from_standard_logging_payload(_ocr_payload([{"index": 0, "markdown": "# Invoice"}]))
+
+    assert data.choices_out == ()
+
+
+def test_ocr_pages_without_markdown_stay_empty():
+    data = LLMCallSpanData.from_standard_logging_payload(
+        _ocr_payload([{"index": 0, "images": []}, "not-a-page"]), capture_content=True
+    )
+
+    assert data.choices_out == ()
+
+
 def test_request_identity_prefers_canonical_team_keys():
     from litellm.integrations.otel.model.payloads import RequestIdentity
 

tests/test_litellm/integrations/otel/test_otel_v2_vendor_mappers.py

@@ -227,6 +227,28 @@ def test_langfuse_mapper_renders_a_responses_api_call_from_the_standard_logging_
     assert attrs["langfuse.observation.type"] == "generation"
 
 
+def test_langfuse_mapper_renders_an_ocr_call_with_the_page_markdown_as_output():
+    payload = {
+        "call_type": "aocr",
+        "custom_llm_provider": "mistral",
+        "model": "mistral-ocr-latest",
+        "messages": None,
+        "response": {
+            "object": "ocr",
+            "model": "mistral-ocr-latest",
+            "pages": [{"index": 0, "markdown": "# Invoice"}, {"index": 1, "markdown": "Total: 42"}],
+            "usage_info": {"pages_processed": 2},
+        },
+    }
+    data = LLMCallSpanData.from_standard_logging_payload(payload, capture_content=True)
+    attrs = LangfuseMapper().map(data)
+
+    assert json.loads(attrs["langfuse.observation.output"]) == [
+        {"role": "assistant", "content": "# Invoice\n\nTotal: 42", "refusal": None, "tool_calls": None}
+    ]
+    assert attrs["langfuse.observation.type"] == "generation"
+
+
 # --------------------------------------------------------------------------- #
 #  Weave
 # --------------------------------------------------------------------------- #