Add LangChain instrumentation with OTel GenAI semantic convention sup…

logfire-api/logfire_api/__init__.py

@@ -191,6 +191,9 @@ def instrument_google_genai(self, *args, **kwargs) -> None: ...
 
             def instrument_litellm(self, *args, **kwargs) -> None: ...
 
+            def instrument_langchain(self, *args, **kwargs) -> ContextManager[None]:
+                return nullcontext()
+
             def instrument_aiohttp_client(self, *args, **kwargs) -> None: ...
 
             def instrument_aiohttp_server(self, *args, **kwargs) -> None: ...
@@ -229,6 +232,7 @@ def shutdown(self, *args, **kwargs) -> None: ...
         instrument_anthropic = DEFAULT_LOGFIRE_INSTANCE.instrument_anthropic
         instrument_google_genai = DEFAULT_LOGFIRE_INSTANCE.instrument_google_genai
         instrument_litellm = DEFAULT_LOGFIRE_INSTANCE.instrument_litellm
+        instrument_langchain = DEFAULT_LOGFIRE_INSTANCE.instrument_langchain
         instrument_asyncpg = DEFAULT_LOGFIRE_INSTANCE.instrument_asyncpg
         instrument_print = DEFAULT_LOGFIRE_INSTANCE.instrument_print
         instrument_celery = DEFAULT_LOGFIRE_INSTANCE.instrument_celery

logfire-api/logfire_api/__init__.pyi

@@ -33,6 +33,7 @@ instrument_openai_agents = DEFAULT_LOGFIRE_INSTANCE.instrument_openai_agents
 instrument_anthropic = DEFAULT_LOGFIRE_INSTANCE.instrument_anthropic
 instrument_google_genai = DEFAULT_LOGFIRE_INSTANCE.instrument_google_genai
 instrument_litellm = DEFAULT_LOGFIRE_INSTANCE.instrument_litellm
+instrument_langchain = DEFAULT_LOGFIRE_INSTANCE.instrument_langchain
 instrument_print = DEFAULT_LOGFIRE_INSTANCE.instrument_print
 instrument_asyncpg = DEFAULT_LOGFIRE_INSTANCE.instrument_asyncpg
 instrument_httpx = DEFAULT_LOGFIRE_INSTANCE.instrument_httpx

logfire-api/logfire_api/_internal/integrations/langchain.pyi

@@ -0,0 +1,4 @@
+import logfire
+from contextlib import AbstractContextManager
+
+def instrument_langchain(logfire_instance: logfire.Logfire) -> AbstractContextManager[None]: ...

logfire-api/logfire_api/_internal/main.pyi

@@ -606,6 +606,17 @@ class Logfire:
         [`openinference-instrumentation-litellm`](https://pypi.org/project/openinference-instrumentation-litellm/)
         package, to which it passes `**kwargs`.
         """
+    def instrument_langchain(self) -> AbstractContextManager[None]:
+        """Instrument LangChain to capture full execution hierarchy with tool definitions.
+
+        This patches LangChain's BaseCallbackManager to inject a callback handler
+        that captures the complete execution hierarchy including chains, tools,
+        retrievers, and LLM calls with tool definitions.
+
+        Returns:
+            A context manager that will revert the instrumentation when exited.
+                Use of this context manager is optional.
+        """
     def instrument_print(self) -> AbstractContextManager[None]:
         """Instrument the built-in `print` function so that calls to it are logged.
 

logfire/__init__.py

@@ -37,6 +37,7 @@
 instrument_anthropic = DEFAULT_LOGFIRE_INSTANCE.instrument_anthropic
 instrument_google_genai = DEFAULT_LOGFIRE_INSTANCE.instrument_google_genai
 instrument_litellm = DEFAULT_LOGFIRE_INSTANCE.instrument_litellm
+instrument_langchain = DEFAULT_LOGFIRE_INSTANCE.instrument_langchain
 instrument_print = DEFAULT_LOGFIRE_INSTANCE.instrument_print
 instrument_asyncpg = DEFAULT_LOGFIRE_INSTANCE.instrument_asyncpg
 instrument_httpx = DEFAULT_LOGFIRE_INSTANCE.instrument_httpx
@@ -132,6 +133,7 @@ def loguru_handler() -> Any:
     'instrument_anthropic',
     'instrument_google_genai',
     'instrument_litellm',
+    'instrument_langchain',
     'instrument_print',
     'instrument_asyncpg',
     'instrument_httpx',

logfire/_internal/exporters/processor_wrapper.py

@@ -328,6 +328,64 @@ def _tweak_fastapi_span(span: ReadableSpanDict):
     span['events'] = new_events[::-1]
 
 
+def _normalize_content_block(block: dict[str, Any]) -> dict[str, Any]:
+    """Normalize a content block to OTel GenAI schema.
+
+    Handles:
+    - Text: converts 'text' field to 'content' (OTel uses 'content')
+    - tool_use: converts to 'tool_call' (OTel standard)
+    - tool_result: converts to 'tool_call_response' (OTel standard)
+    """
+    block_type = block.get('type', 'text')
+
+    if block_type == 'text':
+        return {
+            'type': 'text',
+            'content': block.get('content', block.get('text', '')),
+        }
+
+    if block_type == 'tool_use':
+        return {
+            'type': 'tool_call',
+            'id': block.get('id'),
+            'name': block.get('name'),
+            'arguments': block.get('input', block.get('arguments')),
+        }
+
+    if block_type == 'tool_result':
+        return {
+            'type': 'tool_call_response',
+            'id': block.get('tool_use_id', block.get('id')),
+            'response': block.get('content', block.get('response')),
+        }
+
+    return block
+
+
+def _convert_to_otel_message(msg: dict[str, Any]) -> dict[str, Any]:
+    """Convert a message dict to OTel GenAI message schema with role and parts."""
+    otel_msg: dict[str, Any] = {'role': msg.get('role', 'user'), 'parts': []}
+    content = msg.get('content')
+    if content:
+        if isinstance(content, str):
+            otel_msg['parts'].append({'type': 'text', 'content': content})
+        elif isinstance(content, list):
+            for block in cast(list[Any], content):
+                if isinstance(block, dict):
+                    otel_msg['parts'].append(_normalize_content_block(cast('dict[str, Any]', block)))
+    if tool_calls := msg.get('tool_calls'):
+        for tc in tool_calls:
+            otel_msg['parts'].append(
+                {
+                    'type': 'tool_call',
+                    'id': tc.get('id'),
+                    'name': tc.get('function', {}).get('name') or tc.get('name'),
+                    'arguments': tc.get('function', {}).get('arguments') or tc.get('args'),
+                }
+            )
+    return otel_msg
+
+
 def _transform_langchain_span(span: ReadableSpanDict):
     """Transform spans generated by LangSmith to work better in the Logfire UI.
 
@@ -387,6 +445,19 @@ def _transform_langchain_span(span: ReadableSpanDict):
         # Remove gen_ai.system=langchain as this also interferes with costs in the UI.
         attributes = {k: v for k, v in attributes.items() if k != 'gen_ai.system'}
 
+    # Extract finish reason from completion data
+    with suppress(Exception):
+        completion = parsed_attributes.get('gen_ai.completion', {})
+        stop_reason = (
+            completion.get('generations', [[{}]])[0][0]
+            .get('message', {})
+            .get('kwargs', {})
+            .get('response_metadata', {})
+            .get('stop_reason')
+        )
+        if stop_reason:
+            new_attributes['gen_ai.response.finish_reasons'] = json.dumps([stop_reason])
+
     # Add `all_messages_events`
     with suppress(Exception):
         input_messages = parsed_attributes.get('input.value', parsed_attributes.get('gen_ai.prompt', {}))['messages']
@@ -422,6 +493,33 @@ def _transform_langchain_span(span: ReadableSpanDict):
         new_attributes['all_messages_events'] = json.dumps(message_events)
         properties['all_messages_events'] = {'type': 'array'}
 
+        # Extract OTel GenAI formatted messages
+        input_msgs: list[dict[str, Any]] = []
+        output_msgs: list[dict[str, Any]] = []
+        system_instructions: list[Any] = []
+        for msg in message_events:
+            role = msg.get('role')
+            if role == 'system':
+                content = msg.get('content', '')
+                if isinstance(content, str):
+                    system_instructions.append({'type': 'text', 'content': content})
+                elif isinstance(content, list):
+                    system_instructions.extend(cast(list[Any], content))
+            elif role == 'assistant':
+                output_msgs.append(_convert_to_otel_message(msg))
+            else:
+                input_msgs.append(_convert_to_otel_message(msg))
+
+        if input_msgs:
+            new_attributes['gen_ai.input.messages'] = json.dumps(input_msgs)
+            properties['gen_ai.input.messages'] = {'type': 'array'}
+        if output_msgs:
+            new_attributes['gen_ai.output.messages'] = json.dumps(output_msgs)
+            properties['gen_ai.output.messages'] = {'type': 'array'}
+        if system_instructions:
+            new_attributes['gen_ai.system_instructions'] = json.dumps(system_instructions)
+            properties['gen_ai.system_instructions'] = {'type': 'array'}
+
     span['attributes'] = {
         **attributes,
         ATTRIBUTES_JSON_SCHEMA_KEY: attributes_json_schema(properties),