Skip to content

feat(litellm): Gate prompt/response collection on data_collection option #7180

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ericapisani wants to merge 1 commit into
base: master
Choose a base branch
from
+275 −6
Open

feat(litellm): Gate prompt/response collection on data_collection option #7180

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 17 additions & 6 deletions sentry_sdk/integrations/litellm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,7 +18,7 @@
has_span_streaming_enabled,
should_truncate_gen_ai_input,
)
from sentry_sdk.utils import event_from_exception
from sentry_sdk.utils import event_from_exception, has_data_collection_enabled

if TYPE_CHECKING:
from datetime import datetime
Expand Down Expand Up @@ -129,7 +129,13 @@ def _input_callback(kwargs: "Dict[str, Any]") -> None:
set_data_normalized(span, SPANDATA.GEN_AI_OPERATION_NAME, operation)

# Record input/messages if allowed
if should_send_default_pii() and integration.include_prompts:
record_inputs = False
if has_data_collection_enabled(client.options):
record_inputs = client.options["data_collection"]["gen_ai"]["inputs"]
elif should_send_default_pii() and integration.include_prompts:
record_inputs = True

if record_inputs:
if operation == "embeddings":
# For embeddings, look for the 'input' parameter
embedding_input = kwargs.get("input")
Expand All @@ -141,7 +147,6 @@ def _input_callback(kwargs: "Dict[str, Any]") -> None:
if isinstance(embedding_input, list)
else [embedding_input]
)
client = sentry_sdk.get_client()
messages_data = (
truncate_and_annotate_embedding_inputs(input_list, span, scope)
if should_truncate_gen_ai_input(client.options)
Expand All @@ -158,7 +163,6 @@ def _input_callback(kwargs: "Dict[str, Any]") -> None:
# For chat, look for the 'messages' parameter
messages = kwargs.get("messages", [])
if messages:
client = sentry_sdk.get_client()
scope = sentry_sdk.get_current_scope()
messages = _convert_message_parts(messages)
messages_data = (
Expand Down Expand Up @@ -206,7 +210,8 @@ def _success_callback(
if span is None:
return

integration = sentry_sdk.get_client().get_integration(LiteLLMIntegration)
client = sentry_sdk.get_client()
integration = client.get_integration(LiteLLMIntegration)
if integration is None:
return

Expand All @@ -218,7 +223,13 @@ def _success_callback(
)

# Record response content if allowed
if should_send_default_pii() and integration.include_prompts:
record_outputs = False
if has_data_collection_enabled(client.options):
record_outputs = client.options["data_collection"]["gen_ai"]["outputs"]
elif should_send_default_pii() and integration.include_prompts:
record_outputs = True

if record_outputs:
if hasattr(completion_response, "choices"):
response_messages = []
for choice in completion_response.choices:
Expand Down
258 changes: 258 additions & 0 deletions tests/integrations/litellm/test_litellm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3589,3 +3589,261 @@ def test_convert_message_parts_image_url_missing_url():
converted = _convert_message_parts(messages)
# Should return item unchanged
assert converted[0]["content"][0]["type"] == "image_url"


@pytest.mark.parametrize(
"data_collection, send_default_pii, include_prompts, expected_present, expected_absent",
[
pytest.param(
{"gen_ai": {"inputs": True, "outputs": True}},
False,
False,
[SPANDATA.GEN_AI_REQUEST_MESSAGES, SPANDATA.GEN_AI_RESPONSE_TEXT],
[],
id="gen-ai-inputs-and-outputs-enabled-override-legacy-off",
),
pytest.param(
{"gen_ai": {"inputs": False, "outputs": False}},
True,
True,
[],
[SPANDATA.GEN_AI_REQUEST_MESSAGES, SPANDATA.GEN_AI_RESPONSE_TEXT],
id="gen-ai-inputs-and-outputs-disabled-override-legacy-on",
),
pytest.param(
{"gen_ai": {"inputs": True, "outputs": False}},
False,
False,
[SPANDATA.GEN_AI_REQUEST_MESSAGES],
[SPANDATA.GEN_AI_RESPONSE_TEXT],
id="gen-ai-inputs-enabled-outputs-disabled",
),
pytest.param(
{"gen_ai": {"inputs": False, "outputs": True}},
False,
False,
[SPANDATA.GEN_AI_RESPONSE_TEXT],
[SPANDATA.GEN_AI_REQUEST_MESSAGES],
id="gen-ai-outputs-enabled-inputs-disabled",
),
pytest.param(
{"gen_ai": {}},
False,
False,
[SPANDATA.GEN_AI_REQUEST_MESSAGES, SPANDATA.GEN_AI_RESPONSE_TEXT],
[],
id="gen-ai-inputs-and-outputs-omitted-default-to-enabled",
),
pytest.param(
None,
True,
True,
[SPANDATA.GEN_AI_REQUEST_MESSAGES, SPANDATA.GEN_AI_RESPONSE_TEXT],
[],
id="no-gen-ai-config-legacy-pii-and-include-prompts-enabled",
),
pytest.param(
None,
False,
True,
[],
[SPANDATA.GEN_AI_REQUEST_MESSAGES, SPANDATA.GEN_AI_RESPONSE_TEXT],
id="no-gen-ai-config-legacy-pii-disabled",
),
],
)
def test_chat_completion_data_collection(
reset_litellm_executor,
sentry_init,
capture_events,
get_model_response,
nonstreaming_chat_completions_model_response,
data_collection,
send_default_pii,
include_prompts,
expected_present,
expected_absent,
):
sentry_init_kwargs = dict(
integrations=[LiteLLMIntegration(include_prompts=include_prompts)],
disabled_integrations=[StdlibIntegration],
traces_sample_rate=1.0,
send_default_pii=send_default_pii,
stream_gen_ai_spans=False,
trace_lifecycle="static",
)
if data_collection is not None:
sentry_init_kwargs["_experiments"] = {"data_collection": data_collection}

sentry_init(**sentry_init_kwargs)

messages = [{"role": "user", "content": "Hello!"}]

client = OpenAI(api_key="test-key")

model_response = get_model_response(
nonstreaming_chat_completions_model_response(
response_id="chatcmpl-test",
response_model="gpt-3.5-turbo",
message_content="Test response",
created=1234567890,
usage=CompletionUsage(
prompt_tokens=10,
completion_tokens=20,
total_tokens=30,
),
),
serialize_pydantic=True,
request_headers={"X-Stainless-Raw-Response": "true"},
)

events = capture_events()

with mock.patch.object(
client.completions._client._client,
"send",
return_value=model_response,
), start_transaction(name="litellm test"):
litellm.completion(
model="gpt-3.5-turbo",
messages=messages,
client=client,
)

litellm_utils.executor.shutdown(wait=True)

(event,) = events
(span,) = [
x
for x in event["spans"]
if x["op"] == OP.GEN_AI_CHAT and x["origin"] == "auto.ai.litellm"
]
span_data = span["data"]

for key in expected_present:
assert key in span_data, f"{key} should have been collected"

if SPANDATA.GEN_AI_REQUEST_MESSAGES in expected_present:
assert json.loads(span_data[SPANDATA.GEN_AI_REQUEST_MESSAGES]) == messages

if SPANDATA.GEN_AI_RESPONSE_TEXT in expected_present:
response_text = json.loads(span_data[SPANDATA.GEN_AI_RESPONSE_TEXT])
assert response_text["role"] == "assistant"
assert response_text["content"] == "Test response"

for key in expected_absent:
assert key not in span_data, f"{key} should not have been collected"

# Data collection never gates non-PII attributes
assert span_data[SPANDATA.GEN_AI_OPERATION_NAME] == "chat"
assert span_data[SPANDATA.GEN_AI_REQUEST_MODEL] == "gpt-3.5-turbo"
assert span_data[SPANDATA.GEN_AI_RESPONSE_MODEL] == "gpt-3.5-turbo"
assert span_data[SPANDATA.GEN_AI_USAGE_TOTAL_TOKENS] == 30


@pytest.mark.parametrize(
"data_collection, send_default_pii, include_prompts, should_collect_input",
[
pytest.param(
{"gen_ai": {"inputs": True}},
False,
False,
True,
id="gen-ai-inputs-enabled-override-legacy-off",
),
pytest.param(
{"gen_ai": {"inputs": False}},
True,
True,
False,
id="gen-ai-inputs-disabled-override-legacy-on",
),
pytest.param(
{"gen_ai": {}},
False,
False,
True,
id="gen-ai-inputs-omitted-default-to-enabled",
),
pytest.param(
None,
True,
True,
True,
id="no-gen-ai-config-legacy-pii-and-include-prompts-enabled",
),
pytest.param(
None,
False,
True,
False,
id="no-gen-ai-config-legacy-pii-disabled",
),
],
)
def test_embeddings_data_collection(
sentry_init,
capture_events,
get_model_response,
openai_embedding_model_response,
data_collection,
send_default_pii,
include_prompts,
should_collect_input,
):
sentry_init_kwargs = dict(
integrations=[LiteLLMIntegration(include_prompts=include_prompts)],
disabled_integrations=[StdlibIntegration],
traces_sample_rate=1.0,
send_default_pii=send_default_pii,
stream_gen_ai_spans=False,
trace_lifecycle="static",
)
if data_collection is not None:
sentry_init_kwargs["_experiments"] = {"data_collection": data_collection}

sentry_init(**sentry_init_kwargs)

client = OpenAI(api_key="test-key")

model_response = get_model_response(
openai_embedding_model_response,
serialize_pydantic=True,
request_headers={"X-Stainless-Raw-Response": "true"},
)

events = capture_events()

with mock.patch.object(
client.embeddings._client._client,
"send",
return_value=model_response,
):
with start_transaction(name="litellm test"):
litellm.embedding(
model="text-embedding-ada-002",
input="Hello, world!",
client=client,
)
# Allow time for callbacks to complete (they may run in separate threads)
time.sleep(0.1)

(event,) = events
(span,) = [
x
for x in event["spans"]
if x["op"] == OP.GEN_AI_EMBEDDINGS and x["origin"] == "auto.ai.litellm"
]
span_data = span["data"]

if should_collect_input:
assert json.loads(span_data[SPANDATA.GEN_AI_EMBEDDINGS_INPUT]) == [
"Hello, world!"
]
else:
assert SPANDATA.GEN_AI_EMBEDDINGS_INPUT not in span_data

# Data collection never gates non-PII attributes
assert span_data[SPANDATA.GEN_AI_OPERATION_NAME] == "embeddings"
assert span_data[SPANDATA.GEN_AI_REQUEST_MODEL] == "text-embedding-ada-002"
assert span_data[SPANDATA.GEN_AI_USAGE_INPUT_TOKENS] == 5
Loading