Skip to content

Commit b0fb33f

Browse files
authored
fix(perplexity): map Responses API cache tokens into input_token_details (#39923)
1 parent 502b2b4 commit b0fb33f

2 files changed

Lines changed: 134 additions & 0 deletions

File tree

‎libs/partners/perplexity/langchain_perplexity/chat_models.py‎

Lines changed: 17 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -38,6 +38,7 @@
3838
ToolMessageChunk,
3939
)
4040
from langchain_core.messages.ai import (
41+
InputTokenDetails,
4142
OutputTokenDetails,
4243
UsageMetadata,
4344
subtract_usage,
@@ -344,6 +345,10 @@ def _convert_responses_usage(usage: Any) -> UsageMetadata | None:
344345
Returns `None` if `usage` itself is missing or if either token field is
345346
absent — emitting zeroed `UsageMetadata` would silently undercount usage
346347
in downstream cost dashboards.
348+
349+
Cache hits and cache writes reported under `input_tokens_details` are mapped
350+
onto the standard `InputTokenDetails` slots so downstream consumers can tell
351+
cached input from fresh input.
347352
"""
348353
if usage is None:
349354
return None
@@ -354,10 +359,22 @@ def _convert_responses_usage(usage: Any) -> UsageMetadata | None:
354359
total_tokens = _get_attr(usage, "total_tokens", None)
355360
if total_tokens is None:
356361
total_tokens = input_tokens + output_tokens
362+
363+
input_token_details: InputTokenDetails = {}
364+
details = _get_attr(usage, "input_tokens_details", None)
365+
if details is not None:
366+
cache_read = _get_attr(details, "cache_read_input_tokens", None)
367+
if cache_read is not None:
368+
input_token_details["cache_read"] = cache_read
369+
cache_creation = _get_attr(details, "cache_creation_input_tokens", None)
370+
if cache_creation is not None:
371+
input_token_details["cache_creation"] = cache_creation
372+
357373
return UsageMetadata(
358374
input_tokens=input_tokens,
359375
output_tokens=output_tokens,
360376
total_tokens=total_tokens,
377+
input_token_details=input_token_details,
361378
)
362379

363380

‎libs/partners/perplexity/tests/unit_tests/test_chat_models_responses.py‎

Lines changed: 117 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -850,6 +850,123 @@ def test_convert_responses_usage_derives_total_when_absent() -> None:
850850
assert result["total_tokens"] == 12
851851

852852

853+
def test_convert_responses_usage_maps_cache_token_details() -> None:
854+
details = _make_response_obj(
855+
cache_read_input_tokens=800, cache_creation_input_tokens=150
856+
)
857+
usage = _make_response_obj(
858+
input_tokens=1000,
859+
output_tokens=50,
860+
total_tokens=1050,
861+
input_tokens_details=details,
862+
)
863+
result = _convert_responses_usage(usage)
864+
assert result is not None
865+
assert result["input_token_details"] == {"cache_read": 800, "cache_creation": 150}
866+
867+
868+
def test_convert_responses_usage_maps_cache_token_details_from_mapping() -> None:
869+
"""`_get_attr` accepts dict payloads, so a mapping must map identically."""
870+
usage = {
871+
"input_tokens": 1000,
872+
"output_tokens": 50,
873+
"total_tokens": 1050,
874+
"input_tokens_details": {
875+
"cache_read_input_tokens": 800,
876+
"cache_creation_input_tokens": 150,
877+
},
878+
}
879+
result = _convert_responses_usage(usage)
880+
assert result is not None
881+
assert result["input_token_details"] == {"cache_read": 800, "cache_creation": 150}
882+
883+
884+
def test_convert_responses_usage_maps_only_reported_cache_fields() -> None:
885+
details = _make_response_obj(
886+
cache_read_input_tokens=800, cache_creation_input_tokens=None
887+
)
888+
usage = _make_response_obj(
889+
input_tokens=1000,
890+
output_tokens=50,
891+
total_tokens=1050,
892+
input_tokens_details=details,
893+
)
894+
result = _convert_responses_usage(usage)
895+
assert result is not None
896+
assert result["input_token_details"] == {"cache_read": 800}
897+
898+
899+
def test_convert_responses_usage_omits_cache_details_when_not_reported() -> None:
900+
usage = _make_response_obj(input_tokens=10, output_tokens=2, total_tokens=12)
901+
result = _convert_responses_usage(usage)
902+
assert result is not None
903+
assert result["input_token_details"] == {}
904+
905+
906+
def test_convert_responses_to_chat_result_includes_cache_token_details() -> None:
907+
details = _make_response_obj(
908+
cache_read_input_tokens=800, cache_creation_input_tokens=150
909+
)
910+
usage = _make_response_obj(
911+
input_tokens=1000,
912+
output_tokens=50,
913+
total_tokens=1050,
914+
input_tokens_details=details,
915+
)
916+
response = _make_response_obj(
917+
id="resp_cache",
918+
model="sonar-pro",
919+
status="completed",
920+
object="response",
921+
output_text="hi",
922+
output=[],
923+
usage=usage,
924+
citations=None,
925+
images=None,
926+
related_questions=None,
927+
search_results=None,
928+
)
929+
930+
result = _convert_responses_to_chat_result(response)
931+
message = result.generations[0].message
932+
933+
assert isinstance(message, AIMessage)
934+
assert message.usage_metadata is not None
935+
assert message.usage_metadata["input_token_details"] == {
936+
"cache_read": 800,
937+
"cache_creation": 150,
938+
}
939+
940+
941+
def test_stream_event_conversion_for_completed_includes_cache_token_details() -> None:
942+
details = _make_response_obj(
943+
cache_read_input_tokens=800, cache_creation_input_tokens=150
944+
)
945+
usage = _make_response_obj(
946+
input_tokens=1000,
947+
output_tokens=50,
948+
total_tokens=1050,
949+
input_tokens_details=details,
950+
)
951+
response = _make_response_obj(
952+
id="resp_cache_stream",
953+
model="sonar-pro",
954+
status="completed",
955+
object="response",
956+
usage=usage,
957+
)
958+
chunk = _convert_responses_stream_event_to_chunk(
959+
_make_event("response.completed", response=response)
960+
)
961+
assert chunk is not None
962+
assert isinstance(chunk.message, AIMessageChunk)
963+
assert chunk.message.usage_metadata is not None
964+
assert chunk.message.usage_metadata["input_token_details"] == {
965+
"cache_read": 800,
966+
"cache_creation": 150,
967+
}
968+
969+
853970
# ---------------------------------------------------------------------------
854971
# Error and edge cases in conversion / streaming
855972
# ---------------------------------------------------------------------------

0 commit comments

Comments
 (0)