fix(perplexity): keep citations and search_results on assembled streaming responses

Every Perplexity stream chunk carries the top-level citations and search_results arrays, and a non-streaming response exposes both, but the response rebuilt from a stream dropped them. Success callbacks and spend logs for streamed Perplexity calls therefore never saw the sources. Copy the last streamed values onto the assembled response through the provider streaming metadata hook

litellm/llms/perplexity/chat/transformation.py

@@ -2,12 +2,12 @@
 Translate from OpenAI's `/v1/chat/completions` to Perplexity's `/v1/chat/completions`
 """
 
-from collections.abc import Mapping
+from collections.abc import Mapping, Sequence
 from types import MappingProxyType
 from typing import TYPE_CHECKING, Final
 
 import httpx
-from pydantic import ConfigDict, TypeAdapter
+from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError
 
 import litellm
 from litellm._logging import verbose_logger
@@ -24,6 +24,19 @@ _PERPLEXITY_ATTRIBUTION_HEADERS: Final[Mapping[str, str]] = MappingProxyType({"X
 _JSON_OBJECT: Final = TypeAdapter(dict[str, object], config=ConfigDict(strict=True, hide_input_in_errors=True))
 
 
+class _StreamedSearchSources(BaseModel):
+    model_config = ConfigDict(frozen=True, from_attributes=True)
+    citations: tuple[str, ...] | None = None
+    search_results: tuple[Mapping[str, object], ...] | None = None
+
+
+def _streamed_search_sources(chunk: object) -> _StreamedSearchSources:
+    try:
+        return _StreamedSearchSources.model_validate(chunk)
+    except ValidationError:
+        return _StreamedSearchSources()
+
+
 class PerplexityChatConfig(OpenAIGPTConfig):
     @property
     def custom_llm_provider(self) -> str | None:
@@ -120,6 +133,18 @@ class PerplexityChatConfig(OpenAIGPTConfig):
 
         return model_response
 
+    def apply_assembled_streaming_response_metadata(self, response: ModelResponse, chunks: Sequence[object]) -> None:
+        sources: Final = tuple(map(_streamed_search_sources, reversed(chunks)))
+        citations: Final = next((source.citations for source in sources if source.citations), ())
+        search_results: Final = next((source.search_results for source in sources if source.search_results), ())
+        fields: Final = (
+            ("citations", list(citations)),
+            ("search_results", [dict(result) for result in search_results]),
+        )
+        for field_name, value in fields:
+            if value:
+                setattr(response, field_name, value)
+
     def _enhance_usage_with_perplexity_fields(self, model_response: ModelResponse, raw_response_json: dict) -> None:
         """
         Extract citation tokens and search queries from Perplexity API response

tests/unit/llms/perplexity/chat/test_perplexity_chat_transformation.py

@@ -13,10 +13,10 @@ import pytest
 
 # Add the project root to Python path
 
-from litellm import ModelResponse
+from litellm import ModelResponse, stream_chunk_builder
 from litellm.litellm_core_utils.litellm_logging import Logging
 from litellm.llms.perplexity.chat.transformation import PerplexityChatConfig
-from litellm.types.utils import Usage
+from litellm.types.utils import ModelResponseStream, Usage
 
 
 class TestPerplexityChatTransformation:
@@ -917,3 +917,59 @@ def test_transform_response_keeps_body_values_of_every_json_type_as_the_provider
     assert response.search_results == search_results
     assert [type(result) for result in response.search_results] == [dict, str, int]
     assert repr(response.usage.citation_tokens) == "18"
+
+
+def _perplexity_stream_logging() -> Logging:
+    logging_obj = Logging(
+        model="sonar",
+        messages=[{"role": "user", "content": "hi"}],
+        stream=True,
+        call_type="completion",
+        start_time=datetime.datetime(2025, 1, 1, tzinfo=datetime.timezone.utc),
+        litellm_call_id="perplexity-stream-test",
+        function_id="perplexity-stream-test",
+    )
+    logging_obj.update_environment_variables(litellm_params={}, optional_params={}, custom_llm_provider="perplexity")
+    return logging_obj
+
+
+def _perplexity_stream_chunk(delta: dict, finish_reason: str | None, sources: dict) -> ModelResponseStream:
+    return ModelResponseStream(
+        id="cmpl-1",
+        object="chat.completion.chunk",
+        created=1700000000,
+        model="sonar",
+        choices=[{"index": 0, "delta": delta, "finish_reason": finish_reason}],
+        **sources,
+    )
+
+
+@pytest.mark.parametrize("dumped", [False, True])
+def test_stream_chunk_builder_keeps_perplexity_citations_and_search_results(dumped: bool):
+    sources = {
+        "citations": ["https://a.example", "https://b.example"],
+        "search_results": [{"url": "https://a.example", "title": "A", "date": None}],
+    }
+    chunks = [
+        _perplexity_stream_chunk({"role": "assistant", "content": "Paris [1]"}, None, sources),
+        _perplexity_stream_chunk({"content": " is in France [2]."}, "stop", sources),
+    ]
+
+    response = stream_chunk_builder(
+        chunks=[chunk.model_dump() if dumped else chunk for chunk in chunks], logging_obj=_perplexity_stream_logging()
+    )
+
+    assert response.choices[0].message.content == "Paris [1] is in France [2]."
+    assert response.citations == sources["citations"]
+    assert response.search_results == sources["search_results"]
+
+
+def test_stream_chunk_builder_adds_no_perplexity_sources_when_the_stream_had_none():
+    response = stream_chunk_builder(
+        chunks=[_perplexity_stream_chunk({"role": "assistant", "content": "Hi"}, "stop", {})],
+        logging_obj=_perplexity_stream_logging(),
+    )
+
+    assert response.choices[0].message.content == "Hi"
+    assert not hasattr(response, "citations")
+    assert not hasattr(response, "search_results")