diff --git a/litellm/litellm_core_utils/streaming_chunk_builder_utils.py b/litellm/litellm_core_utils/streaming_chunk_builder_utils.py index 68cdc204af4..fe2b129b0e2 100644 --- a/litellm/litellm_core_utils/streaming_chunk_builder_utils.py +++ b/litellm/litellm_core_utils/streaming_chunk_builder_utils.py @@ -706,7 +706,7 @@ class ChunkProcessor: content = delta.get(delta_key, "") if content is None: continue # openai v1.0.0 sets content = None for chunks - content_list.append(content) + content_list.append(delta_content_as_str(content)) # Combine the "content" strings into a single string || combine the 'function' strings into a single string combined_content: Final = "".join(content_list) @@ -1476,12 +1476,33 @@ def _delta_carries_more_than_text(delta: Mapping[str, object]) -> bool: return any(delta.get(field) is not None for field in _NON_TEXT_DELTA_FIELDS) +def delta_content_as_str(content: object) -> str: + """Coerce a streamed delta content value to plain text. + + OpenAI-compatible providers are supposed to stream delta.content as a + string, but some stream a list of content parts instead. Join the text + parts into one string and skip every part that carries no usable text, so + the running answer text neither loses the content nor raises when a part + holds a null value or is not a mapping at all. + """ + if isinstance(content, str): + return content + if not isinstance(content, list): + return "" + texts: list[str] = [] + for part in content: + text = part.get("text") if isinstance(part, Mapping) else getattr(part, "text", None) + if isinstance(text, str): + texts.append(text) + return "".join(texts) + + def _simple_text_part(choices: Sequence[object]) -> str | None: deltas: Final = tuple(_stream_choice_delta(choice) for choice in choices) if any(_delta_carries_more_than_text(delta) for delta in deltas): return None content: Final = deltas[0].get("content") - return content if isinstance(content, str) else "" + return delta_content_as_str(content) def stream_chunk_builder( diff --git a/tests/unit/litellm_core_utils/test_streaming_chunk_builder_utils.py b/tests/unit/litellm_core_utils/test_streaming_chunk_builder_utils.py index 5c56245cf33..70796fe1d1d 100644 --- a/tests/unit/litellm_core_utils/test_streaming_chunk_builder_utils.py +++ b/tests/unit/litellm_core_utils/test_streaming_chunk_builder_utils.py @@ -12,7 +12,11 @@ import respx import litellm from litellm import ChatCompletionUsageBlock, stream_chunk_builder from litellm.types.utils import GenericStreamingChunk -from litellm.litellm_core_utils.streaming_chunk_builder_utils import ChunkProcessor, stream_chunk_builder_text_completion +from litellm.litellm_core_utils.streaming_chunk_builder_utils import ( + ChunkProcessor, + delta_content_as_str, + stream_chunk_builder_text_completion, +) from litellm.llms.anthropic.chat.handler import ModelResponseIterator from litellm.types.utils import ( ChatCompletionDeltaToolCall, @@ -2517,3 +2521,65 @@ def test_stream_chunk_builder_keeps_reasoning_items_on_text_only_replies(): assert response is not None assert response.choices[0].message.content == "Hello world" assert response.choices[0].message.reasoning_items == reasoning_items + + +def test_stream_chunk_builder_joins_content_part_list_delta(): + """A provider streaming delta.content as a content-part list must aggregate as text.""" + chunks = [ + ModelResponseStream( + id="chatcmpl-123", + created=1, + model="gpt-4.1-mini", + object="chat.completion.chunk", + choices=[ + StreamingChoices( + finish_reason=None, + index=0, + delta=Delta(content=[{"type": "text", "text": "Hello "}], role="assistant"), + ) + ], + ), + ModelResponseStream( + id="chatcmpl-123", + created=2, + model="gpt-4.1-mini", + object="chat.completion.chunk", + choices=[ + StreamingChoices( + finish_reason="stop", + index=0, + delta=Delta(content=[{"type": "text", "text": "world"}], role=None), + ) + ], + ), + ] + + response = stream_chunk_builder(chunks=chunks) + assert response is not None + assert response.choices[0].message.content == "Hello world" + + +def test_delta_content_as_str_skips_parts_without_usable_text(): + """Parts holding a null value or not being mappings must not abort the join.""" + assert delta_content_as_str([{"type": "text", "text": "a"}, {"type": "text", "text": "b"}]) == "ab" + assert delta_content_as_str([{"type": "text", "text": None}]) == "" + assert delta_content_as_str(["bare string", 42, None]) == "" + assert ( + delta_content_as_str( + [ + {"type": "text", "text": "a"}, + {"type": "text", "text": None}, + {"type": "text", "text": "b"}, + ] + ) + == "ab" + ) + + +def test_delta_content_as_str_reads_text_from_non_mapping_parts(): + """A part exposing only a ``text`` attribute still contributes its text.""" + + class _TextPart: + text = "from attribute" + + assert delta_content_as_str([_TextPart()]) == "from attribute" diff --git a/tests/unit/litellm_core_utils/test_streaming_handler.py b/tests/unit/litellm_core_utils/test_streaming_handler.py index cc1fa1b8803..0ee22a12fac 100644 --- a/tests/unit/litellm_core_utils/test_streaming_handler.py +++ b/tests/unit/litellm_core_utils/test_streaming_handler.py @@ -7844,3 +7844,68 @@ async def test_leaving_an_async_with_block_closes_the_provider_stream(): async with _wrapper_over(stream): pass assert stream.close_calls == 1 + + +def _chunk_with_list_content(text: str, chunk_id: str, created: int) -> ModelResponseStream: + return ModelResponseStream( + id=chunk_id, + object="chat.completion.chunk", + created=created, + model="gpt-5.6", + choices=[ + StreamingChoices( + index=0, + delta=Delta(role="assistant", content=[{"type": "text", "text": text}]), + finish_reason=None, + ) + ], + ) + + +def test_streaming_delta_content_list_accumulates_as_str(): + """A provider streaming delta.content as a content-part list must still accumulate as text.""" + chunks: Final = [ + _chunk_with_list_content("hello ", "chunk1", 1741037890), + _chunk_with_list_content("world", "chunk2", 1741037891), + ] + + response: Final = CustomStreamWrapper( + completion_stream=iter(chunks), + model="gpt-5.6", + logging_obj=MagicMock(), + custom_llm_provider="openai", + ) + + for _ in response: + pass + + assert response.response_uptil_now == "hello world" + + +def test_streaming_accumulates_content_part_with_null_text(): + """A content part carrying a null value must not abort the accumulated stream.""" + chunk: Final = ModelResponseStream( + id="chunk-null", + object="chat.completion.chunk", + created=1741037892, + model="gpt-5.6", + choices=[ + StreamingChoices( + index=0, + delta=Delta(role="assistant", content=[{"type": "text", "text": None}]), + finish_reason=None, + ) + ], + ) + + response: Final = CustomStreamWrapper( + completion_stream=iter([chunk]), + model="gpt-5.6", + logging_obj=MagicMock(), + custom_llm_provider="openai", + ) + + for _ in response: + pass + + assert response.response_uptil_now == ""