fix(ai): add message truncation in langchain

Author: shellmayr

sentry_sdk/integrations/langchain.py

@@ -9,6 +9,7 @@
     normalize_message_roles,
     set_data_normalized,
     get_start_span_function,
+    truncate_and_annotate_messages,
 )
 from sentry_sdk.consts import OP, SPANDATA
 from sentry_sdk.integrations import DidNotEnable, Integration
@@ -221,12 +222,17 @@ def on_llm_start(
                     }
                     for prompt in prompts
                 ]
-                set_data_normalized(
-                    span,
-                    SPANDATA.GEN_AI_REQUEST_MESSAGES,
-                    normalized_messages,
-                    unpack=False,
+                scope = sentry_sdk.get_current_scope()
+                messages_data = truncate_and_annotate_messages(
+                    normalized_messages, span, scope
                 )
+                if messages_data is not None:
+                    set_data_normalized(
+                        span,
+                        SPANDATA.GEN_AI_REQUEST_MESSAGES,
+                        messages_data,
+                        unpack=False,
+                    )
 
     def on_chat_model_start(self, serialized, messages, *, run_id, **kwargs):
         # type: (SentryLangchainCallback, Dict[str, Any], List[List[BaseMessage]], UUID, Any) -> Any
@@ -278,13 +284,17 @@ def on_chat_model_start(self, serialized, messages, *, run_id, **kwargs):
                             self._normalize_langchain_message(message)
                         )
                 normalized_messages = normalize_message_roles(normalized_messages)
-
-                set_data_normalized(
-                    span,
-                    SPANDATA.GEN_AI_REQUEST_MESSAGES,
-                    normalized_messages,
-                    unpack=False,
+                scope = sentry_sdk.get_current_scope()
+                messages_data = truncate_and_annotate_messages(
+                    normalized_messages, span, scope
                 )
+                if messages_data is not None:
+                    set_data_normalized(
+                        span,
+                        SPANDATA.GEN_AI_REQUEST_MESSAGES,
+                        messages_data,
+                        unpack=False,
+                    )
 
     def on_chat_model_end(self, response, *, run_id, **kwargs):
         # type: (SentryLangchainCallback, LLMResult, UUID, Any) -> Any
@@ -758,12 +768,17 @@ def new_invoke(self, *args, **kwargs):
                 and integration.include_prompts
             ):
                 normalized_messages = normalize_message_roles([input])
-                set_data_normalized(
-                    span,
-                    SPANDATA.GEN_AI_REQUEST_MESSAGES,
-                    normalized_messages,
-                    unpack=False,
+                scope = sentry_sdk.get_current_scope()
+                messages_data = truncate_and_annotate_messages(
+                    normalized_messages, span, scope
                 )
+                if messages_data is not None:
+                    set_data_normalized(
+                        span,
+                        SPANDATA.GEN_AI_REQUEST_MESSAGES,
+                        messages_data,
+                        unpack=False,
+                    )
 
             output = result.get("output")
             if (
@@ -813,12 +828,17 @@ def new_stream(self, *args, **kwargs):
             and integration.include_prompts
         ):
             normalized_messages = normalize_message_roles([input])
-            set_data_normalized(
-                span,
-                SPANDATA.GEN_AI_REQUEST_MESSAGES,
-                normalized_messages,
-                unpack=False,
+            scope = sentry_sdk.get_current_scope()
+            messages_data = truncate_and_annotate_messages(
+                normalized_messages, span, scope
             )
+            if messages_data is not None:
+                set_data_normalized(
+                    span,
+                    SPANDATA.GEN_AI_REQUEST_MESSAGES,
+                    messages_data,
+                    unpack=False,
+                )
 
         # Run the agent
         result = f(self, *args, **kwargs)

tests/integrations/langchain/test_langchain.py

@@ -958,3 +958,70 @@ def test_langchain_message_role_normalization_units():
     assert normalized[3]["role"] == "system"  # system unchanged
     assert "role" not in normalized[4]  # Message without role unchanged
     assert normalized[5] == "string message"  # String message unchanged
+
+
+def test_langchain_message_truncation(sentry_init, capture_events):
+    """Test that large messages are truncated properly in Langchain integration."""
+    from langchain_core.outputs import LLMResult, Generation
+
+    sentry_init(
+        integrations=[LangchainIntegration(include_prompts=True)],
+        traces_sample_rate=1.0,
+        send_default_pii=True,
+    )
+    events = capture_events()
+
+    callback = SentryLangchainCallback(max_span_map_size=100, include_prompts=True)
+
+    run_id = "12345678-1234-1234-1234-123456789012"
+    serialized = {"_type": "openai-chat", "model_name": "gpt-3.5-turbo"}
+
+    large_content = (
+        "This is a very long message that will exceed our size limits. " * 1000
+    )
+    prompts = [large_content, large_content, large_content, large_content]
+
+    with start_transaction():
+        callback.on_llm_start(
+            serialized=serialized,
+            prompts=prompts,
+            run_id=run_id,
+            invocation_params={
+                "temperature": 0.7,
+                "max_tokens": 100,
+                "model": "gpt-3.5-turbo",
+            },
+        )
+
+        response = LLMResult(
+            generations=[[Generation(text="The response")]],
+            llm_output={
+                "token_usage": {
+                    "total_tokens": 25,
+                    "prompt_tokens": 10,
+                    "completion_tokens": 15,
+                }
+            },
+        )
+        callback.on_llm_end(response=response, run_id=run_id)
+
+    assert len(events) > 0
+    tx = events[0]
+    assert tx["type"] == "transaction"
+
+    llm_spans = [
+        span for span in tx.get("spans", []) if span.get("op") == "gen_ai.pipeline"
+    ]
+    assert len(llm_spans) > 0
+
+    llm_span = llm_spans[0]
+    assert SPANDATA.GEN_AI_REQUEST_MESSAGES in llm_span["data"]
+
+    messages_data = llm_span["data"][SPANDATA.GEN_AI_REQUEST_MESSAGES]
+    assert isinstance(messages_data, str)
+
+    import json
+
+    parsed_messages = json.loads(messages_data)
+    assert isinstance(parsed_messages, list)
+    assert len(parsed_messages) <= len(prompts)