# LangChain Coralogix's AI Observability integrations make it easy to monitor any LangChain-powered application. With a dedicated LangChain integration, Coralogix consolidates spans emitted by OpenAI, AWS Bedrock, and future LangChain chat providers so teams can understand performance, drift, and tool usage without stitching logs across services. ## How LangChain works This library ships production-ready [OpenTelemetry instrumentation](https://github.com/open-telemetry/opentelemetry-python-contrib/) for LangChain. It automatically hooks into LangChain's callback manager, traces chat and tool-call workflows, and records usage metrics. The integration accelerates debugging, highlights token consumption, and standardizes observability across multi provider LangChain deployments. Instrumentation currently supports [OpenAI](https://coralogix.com/docs/user-guides/ai-observability/setup/openai/index.md) and [AWS Bedrock](https://coralogix.com/docs/user-guides/ai-observability/setup/bedrock/index.md) providers (including tool binding) across synchronous, streaming, and multi turn LangChain runs. ## What you need - Python 3.10–3.13. - Coralogix [API keys](https://coralogix.com/docs/user-guides/account-management/api-keys/api-keys/). - LangChain 1.0.0 or newer along with the provider-specific SDKs (for example `langchain-openai` or `langchain-aws`). > Ensure each underlying provider SDK (OpenAI, AWS Bedrock, etc.) is configured with valid credentials before enabling instrumentation. ## Installation Run the following command. ```text pip install llm-tracekit[langchain] ``` ## Authentication Exporting spans to Coralogix requires configuring OTLP credentials before enabling instrumentation. Use `setup_export_to_coralogix` or the corresponding environment variables to supply authentication details. ### Using `setup_export_to_coralogix` ```text from llm_tracekit import setup_export_to_coralogix setup_export_to_coralogix( coralogix_token="", coralogix_endpoint="ingress.[[DOMAIN_VALUE]]:443", application_name="", subsystem_name="", capture_content=True ) ``` ### Using environment variables If arguments are not passed to `setup_export_to_coralogix`, the helper reads the following environment variables: - `CX_TOKEN`: Your Coralogix API key. - `CX_ENDPOINT`: Select the ingress.\[[DOMAIN_VALUE]\]:443 endpoint that corresponds to your Coralogix [domain](https://coralogix.com/docs/user-guides/account-management/account-settings/coralogix-domain/index.md) using the domain selector at the top of the page. - `CX_APPLICATION_NAME`: Your application's name. - `CX_SUBSYSTEM_NAME`: Your subsystem's name. ## Usage Instrument for LangChain. ## Set up tracing ### Instrument LangChain instrumentation is installed globally by wrapping `BaseCallbackManager`. Create an instance of `LangChainInstrumentor` and call `instrument` before instantiating your chains. ```text from llm_tracekit import LangChainInstrumentor instrumentor = LangChainInstrumentor() instrumentor.instrument() ``` ### Uninstrument To remove instrumentation, call the `uninstrument` method. ```text instrumentor.uninstrument() ``` ### Full example ```py from langchain_openai import ChatOpenAI from llm_tracekit import setup_export_to_coralogix from llm_tracekit import LangChainInstrumentor setup_export_to_coralogix( coralogix_token="", coralogix_endpoint="", application_name="", subsystem_name="" ) instrumentor = LangChainInstrumentor() instrumentor.instrument() llm = ChatOpenAI(model="gpt-4o-mini") response = llm.invoke("Draft a release note for LangChain instrumentation") print(response) ``` ### Enable message content capture By default, message content - such as prompts, completions, tool call arguments, and tool responses - is not captured. To capture message content as span attributes, set the environment variable `OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT` to `true`. Many Coralogix AI evaluations rely on message content; enabling capture is a **best practice**. ### Key differences from OpenTelemetry Prompts, tool invocations, and model responses are stored as span attributes instead of log events, preserving a single correlated timeline for each LangChain run. ## Semantic conventions | **Attribute** | **Type** | **Description** | **Example** | | ------------------------------------------------------------------------------------ | -------- | ---------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------- | | `gen_ai.operation.name` | string | The specific name of the LangChain operation being performed | `chat` | | `gen_ai.system` | string | The provider or framework responsible for the operation | `openai` / `aws.bedrock` | | `gen_ai.request.model` | string | The name of the model requested by the user or application | `gpt-4o-mini` | | `gen_ai.request.temperature` | float | The `temperature` parameter passed in the request. Higher values increase randomness, while lower values make output more deterministic. | `0.2` | | `gen_ai.request.top_p` | float | The `top_p` parameter used for nucleus sampling, limiting token selection to the top probability mass. | `0.95` | | `gen_ai.prompt..role` | string | Role of the author for each input message | `user` | | `gen_ai.prompt..content` | string | Contents of the input message (captured when content capture is enabled) | `Draft a release note for LangChain instrumentation.` | | `gen_ai.prompt..tool_calls..id` | string | ID of a tool call issued from the prompt | `call_yPIxaozNPCSp1tJ34Hsbdtzg` | | `gen_ai.prompt..tool_calls..type` | string | Type of tool call issued from the prompt | `tool_call` | | `gen_ai.prompt..tool_calls..function.name` | string | Function name used in the prompt's tool call | `get_current_weather` | | `gen_ai.prompt..tool_calls..function.arguments` | string | Arguments passed to the prompt's tool call | `{"location": "Paris"}` | | `gen_ai.request.tools..type` | string | Declared type for each available tool definition exposed to the model (provider-agnostic). | `function` | | `gen_ai.request.tools..function.name` | string | Tool name surfaced in the request-level invocation params (OpenAI `function.name`, Bedrock `name`, etc.). | `get_destination_tip` | | `gen_ai.request.tools..function.description` | string | Tool description shown to the LLM regardless of provider. | `Return a mock travel tip for the provided city.` | | `gen_ai.request.tools..function.parameters` | string | JSON-serialized parameter schema (OpenAI `function.parameters`, Bedrock `input_schema`, other provider equivalents). | `{"type":"object","properties":{"city":{"type":"string"}}}` | | `gen_ai.completion..role` | string | Role of the author for each returned choice | `assistant` | | `gen_ai.completion..finish_reason` | string | Finish reason reported for the choice | `stop` | | `gen_ai.completion..content` | string | Text returned by the model for the choice (captured when content capture is enabled) | `Here is the requested release note...` | | `gen_ai.completion..tool_calls..id` | string | ID of a tool call triggered by the model | `call_O8NOz8VlxosSASEsOY7LDUcP` | | `gen_ai.completion..tool_calls..type` | string | Type of tool call triggered by the model | `tool_call` | | `gen_ai.completion..tool_calls..function.name` | string | Function name executed by the model's tool call | `get_current_weather` | | `gen_ai.completion..tool_calls..function.arguments` | string | Arguments supplied to the model's tool call | `{"location": "Paris"}` | | `gen_ai.response.model` | string | Exact model identifier that produced the response | `gpt-4o-mini` | | `gen_ai.response.id` | string | Unique identifier assigned to the specific LangChain response | `run-20241010-abcdef1234567890` | | `gen_ai.response.finish_reasons` | string[] | Ordered list of finish reasons observed across all choices | `["stop"]` | | `gen_ai.usage.input_tokens` | int | Number of tokens consumed by the prompt | `744` | | `gen_ai.usage.output_tokens` | int | Number of tokens generated in the response | `256` |