Skip to content

work on chat span #4696

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

Merged
merged 8 commits into from
Aug 13, 2025
+67 −76
Merged

work on chat span #4696

Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
dd166a8
updated tool span
antonpirker Aug 13, 2025
adaabbf
.
antonpirker Aug 13, 2025
3795d63
tool name
antonpirker Aug 13, 2025
eb95293
chat span
antonpirker Aug 13, 2025
8b4719b
Merge branch 'shellmayr/feat/langchain-integration-update' into anton…
antonpirker Aug 13, 2025
9b09474
fix available tools
antonpirker Aug 13, 2025
94fd28f
cleanup
antonpirker Aug 13, 2025
11ea5ea
Merge branch 'shellmayr/feat/langchain-integration-update' into anton…
antonpirker Aug 13, 2025
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
143 changes: 67 additions & 76 deletions sentry_sdk/integrations/langchain.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,27 +34,20 @@


DATA_FIELDS = {
"temperature": SPANDATA.GEN_AI_REQUEST_TEMPERATURE,
"top_p": SPANDATA.GEN_AI_REQUEST_TOP_P,
"top_k": SPANDATA.GEN_AI_REQUEST_TOP_K,
"frequency_penalty": SPANDATA.GEN_AI_REQUEST_FREQUENCY_PENALTY,
"function_call": SPANDATA.GEN_AI_RESPONSE_TOOL_CALLS,
"tool_calls": SPANDATA.GEN_AI_RESPONSE_TOOL_CALLS,
"tools": SPANDATA.GEN_AI_REQUEST_AVAILABLE_TOOLS,
"response_format": SPANDATA.GEN_AI_RESPONSE_FORMAT,
"logit_bias": SPANDATA.GEN_AI_REQUEST_LOGIT_BIAS,
"max_tokens": SPANDATA.GEN_AI_REQUEST_MAX_TOKENS,
"presence_penalty": SPANDATA.GEN_AI_REQUEST_PRESENCE_PENALTY,
"response_format": SPANDATA.GEN_AI_RESPONSE_FORMAT,
"tags": SPANDATA.GEN_AI_REQUEST_TAGS,
"temperature": SPANDATA.GEN_AI_REQUEST_TEMPERATURE,
"tool_calls": SPANDATA.GEN_AI_RESPONSE_TOOL_CALLS,
"tools": SPANDATA.GEN_AI_REQUEST_AVAILABLE_TOOLS,
"top_k": SPANDATA.GEN_AI_REQUEST_TOP_K,
"top_p": SPANDATA.GEN_AI_REQUEST_TOP_P,
}

# TODO(shellmayr): is this still the case?
# To avoid double collecting tokens, we do *not* measure
# token counts for models for which we have an explicit integration
NO_COLLECT_TOKEN_MODELS = [
# "openai-chat",
# "anthropic-chat",
"cohere-chat",
"huggingface_endpoint",
]


class LangchainIntegration(Integration):
identifier = "langchain"
Expand All @@ -80,7 +73,6 @@ def setup_once():

class WatchedSpan:
span = None # type: Span
no_collect_tokens = False # type: bool
children = [] # type: List[WatchedSpan]
is_pipeline = False # type: bool

Expand Down Expand Up @@ -270,7 +262,7 @@ def on_llm_start(
all_params.update(serialized.get("kwargs", {}))

watched_span = self._create_span(
run_id,
run_id=run_id,
parent_id=parent_run_id,
op=OP.GEN_AI_PIPELINE,
name=kwargs.get("name") or "Langchain LLM call",
Expand All @@ -297,25 +289,31 @@ def on_chat_model_start(self, serialized, messages, *, run_id, **kwargs):
return
all_params = kwargs.get("invocation_params", {})
all_params.update(serialized.get("kwargs", {}))

model = (
all_params.get("model")
or all_params.get("model_name")
or all_params.get("model_id")
or ""
)

watched_span = self._create_span(
run_id,
run_id=run_id,
parent_id=kwargs.get("parent_run_id"),
op=OP.GEN_AI_CHAT,
name=kwargs.get("name") or "Langchain Chat Model",
name=f"chat {model}".strip(),
origin=LangchainIntegration.origin,
)
span = watched_span.span
model = all_params.get(
"model", all_params.get("model_name", all_params.get("model_id"))
)
watched_span.no_collect_tokens = any(
x in all_params.get("_type", "") for x in NO_COLLECT_TOKEN_MODELS
)

if not model and "anthropic" in all_params.get("_type"):
model = "claude-2"
span.set_data(SPANDATA.GEN_AI_OPERATION_NAME, "chat")
if model:
span.set_data(SPANDATA.GEN_AI_REQUEST_MODEL, model)

for key, attribute in DATA_FIELDS.items():
if key in all_params:
set_data_normalized(span, attribute, all_params[key], unpack=False)

Copy link

@cursor cursor bot Aug 13, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bug: Callback Argument Error Causes Telemetry Loss

The on_chat_model_start callback calls set_data_normalized with an unpack=False keyword argument. This likely causes a TypeError that's silently swallowed, preventing request parameters from being recorded on chat spans and degrading telemetry.

Fix in Cursor Fix in Web

if should_send_default_pii() and self.include_prompts:
set_data_normalized(
span,
Expand All @@ -325,18 +323,13 @@ def on_chat_model_start(self, serialized, messages, *, run_id, **kwargs):
for list_ in messages
],
)
for k, v in DATA_FIELDS.items():
if k in all_params:
set_data_normalized(span, v, all_params[k])
# no manual token counting

def on_chat_model_end(self, response, *, run_id, **kwargs):
# type: (SentryLangchainCallback, LLMResult, UUID, Any) -> Any
"""Run when Chat Model ends running."""
with capture_internal_exceptions():
if not run_id:
return

token_usage = None

# Try multiple paths to extract token usage, prioritizing streaming-aware approaches
Expand Down Expand Up @@ -367,27 +360,26 @@ def on_chat_model_end(self, response, *, run_id, **kwargs):
[[x.text for x in list_] for list_ in response.generations],
)

if not span_data.no_collect_tokens:
if token_usage:
input_tokens, output_tokens, total_tokens = (
self._extract_token_usage(token_usage)
)
else:
input_tokens, output_tokens, total_tokens = (
self._extract_token_usage_from_generations(response.generations)
)
if token_usage:
input_tokens, output_tokens, total_tokens = self._extract_token_usage(
token_usage
)
else:
input_tokens, output_tokens, total_tokens = (
self._extract_token_usage_from_generations(response.generations)
)

if (
input_tokens is not None
or output_tokens is not None
or total_tokens is not None
):
record_token_usage(
span_data.span,
input_tokens=input_tokens,
output_tokens=output_tokens,
total_tokens=total_tokens,
)
if (
input_tokens is not None
or output_tokens is not None
or total_tokens is not None
):
record_token_usage(
span_data.span,
input_tokens=input_tokens,
output_tokens=output_tokens,
total_tokens=total_tokens,
)
Copy link

@cursor cursor bot Aug 13, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bug: Token Usage Metrics Inflated Due to Integration Change

The Langchain integration now unconditionally records token usage for both chat and generic LLM spans. The no_collect_tokens guard, which previously prevented double-counting for models with dedicated Sentry integrations, was removed. This leads to inflated token usage metrics in production.

Additional Locations (1)
Fix in Cursor Fix in Web


self._exit_span(span_data, run_id)

Expand Down Expand Up @@ -429,27 +421,26 @@ def on_llm_end(self, response, *, run_id, **kwargs):
[[x.text for x in list_] for list_ in response.generations],
)

if not span_data.no_collect_tokens:
if token_usage:
input_tokens, output_tokens, total_tokens = (
self._extract_token_usage(token_usage)
)
else:
input_tokens, output_tokens, total_tokens = (
self._extract_token_usage_from_generations(response.generations)
)
if token_usage:
input_tokens, output_tokens, total_tokens = self._extract_token_usage(
token_usage
)
else:
input_tokens, output_tokens, total_tokens = (
self._extract_token_usage_from_generations(response.generations)
)

if (
input_tokens is not None
or output_tokens is not None
or total_tokens is not None
):
record_token_usage(
span_data.span,
input_tokens=input_tokens,
output_tokens=output_tokens,
total_tokens=total_tokens,
)
if (
input_tokens is not None
or output_tokens is not None
or total_tokens is not None
):
record_token_usage(
span_data.span,
input_tokens=input_tokens,
output_tokens=output_tokens,
total_tokens=total_tokens,
)

self._exit_span(span_data, run_id)

Expand Down Expand Up @@ -515,13 +506,13 @@ def on_tool_start(self, serialized, input_str, *, run_id, **kwargs):
if not run_id:
return

tool_name = serialized.get("name") or kwargs.get("name")
tool_name = serialized.get("name") or kwargs.get("name") or ""

watched_span = self._create_span(
run_id,
run_id=run_id,
parent_id=kwargs.get("parent_run_id"),
op=OP.GEN_AI_EXECUTE_TOOL,
name=f"execute_tool {tool_name}",
name=f"execute_tool {tool_name}".strip(),
origin=LangchainIntegration.origin,
)
span = watched_span.span
Expand Down
Loading