Overview
Monitor, troubleshoot, and evaluate your LLM-powered applications (e.g. chatbot, data extraction tool, etc) built using LangChain.
Use LLM Observability to investigate the root cause of issues, monitor operational performance, and evaluate the quality, privacy, and safety of your LLM applications.
See the LLM Observability tracing view video for an example of how you can investigate a trace.
Get cost estimation, prompt and completion sampling, error tracking, performance metrics, and more out of LangChain Python library requests using Datadog metrics, APM, and logs.
Setup
LLM Observability: Get end-to-end visibility into your LLM application using LangChain
You can enable LLM Observability in different environments. Follow the appropriate setup based on your scenario:
Installation for Python
If you do not have the Datadog Agent:
Install the ddtrace
package:
Start your application with the following command, enabling Agentless mode:
DD_SITE=<YOUR_DATADOG_SITE> DD_API_KEY=<YOUR_API_KEY> DD_LLMOBS_ENABLED=1 DD_LLMOBS_AGENTLESS_ENABLED=1 DD_LLMOBS_ML_APP=<YOUR_ML_APP_NAME> ddtrace-run python <YOUR_APP>.py
If you already have the Datadog Agent installed:
Make sure the Agent is running and that APM and StatsD are enabled. For example, use the following command with Docker:
docker run -d \
--cgroupns host \
--pid host \
-v /var/run/docker.sock:/var/run/docker.sock:ro \
-v /proc/:/host/proc/:ro \
-v /sys/fs/cgroup/:/host/sys/fs/cgroup:ro \
-e DD_API_KEY=<DATADOG_API_KEY> \
-p 127.0.0.1:8126:8126/tcp \
-p 127.0.0.1:8125:8125/udp \
-e DD_DOGSTATSD_NON_LOCAL_TRAFFIC=true \
-e DD_APM_ENABLED=true \
gcr.io/datadoghq/agent:latest
Install the ddtrace
package if it isn’t installed yet:
Start your application using the ddtrace-run
command to automatically enable tracing:
DD_SITE=<YOUR_DATADOG_SITE> DD_API_KEY=<YOUR_API_KEY> DD_LLMOBS_ENABLED=1 DD_LLMOBS_ML_APP=<YOUR_ML_APP_NAME> ddtrace-run python <YOUR_APP>.py
Note: If the Agent is running on a custom host or port, set DD_AGENT_HOST
and DD_TRACE_AGENT_PORT
accordingly.
If you are running LLM Observability in a serverless environment (AWS Lambda):
Install the Datadog-Python and Datadog-Extension Lambda layers as part of your AWS Lambda setup.
Enable LLM Observability by setting the following environment variables:
DD_SITE=<YOUR_DATADOG_SITE> DD_API_KEY=<YOUR_API_KEY> DD_LLMOBS_ENABLED=1 DD_LLMOBS_ML_APP=<YOUR_ML_APP_NAME>
Note: In serverless environments, Datadog automatically flushes spans when the Lambda function finishes running.
Automatic LangChain Tracing
LangChain integration is automatically enabled when LLM Observability is configured. This captures latency, errors, input/output messages, and token usage for LangChain operations.
The following methods are traced for both synchronous and asynchronous LangChain operations:
- LLMs:
llm.invoke()
, llm.ainvoke()
- Chat Models:
chat_model.invoke()
, chat_model.ainvoke()
- Chains/LCEL:
chain.invoke()
, chain.ainvoke()
, chain.batch()
, chain.abatch()
- Embeddings:
OpenAIEmbeddings.embed_documents()
, OpenAIEmbeddings.embed_query()
No additional setup is required for these methods.
Validation
Validate that LLM Observability is properly capturing spans by checking your application logs for successful span creation. You can also run the following command to check the status of the ddtrace
integration:
Look for the following message to confirm the setup:
Debugging
If you encounter issues during setup, enable debug logging by passing the --debug
flag:
This will display any errors related to data transmission or instrumentation, including issues with LangChain traces.
APM: Get Usage Metrics for your python Applications
Enable APM and StatsD in your Datadog Agent. For example, in Docker:
docker run -d --cgroupns host \
--pid host \
-v /var/run/docker.sock:/var/run/docker.sock:ro \
-v /proc/:/host/proc/:ro \
-v /sys/fs/cgroup/:/host/sys/fs/cgroup:ro \
-e DD_API_KEY=<DATADOG_API_KEY> \
-p 127.0.0.1:8126:8126/tcp \
-p 127.0.0.1:8125:8125/udp \
-e DD_DOGSTATSD_NON_LOCAL_TRAFFIC=true \
-e DD_APM_ENABLED=true \
gcr.io/datadoghq/agent:latest
Install the Datadog APM Python library.
pip install ddtrace>=1.17
Prefix your LangChain Python application command with ddtrace-run
.
DD_SERVICE="my-service" DD_ENV="staging" DD_API_KEY=<DATADOG_API_KEY> ddtrace-run python <your-app>.py
Note: If the Agent is using a non-default hostname or port, be sure to also set DD_AGENT_HOST
, DD_TRACE_AGENT_PORT
, or DD_DOGSTATSD_PORT
.
See the APM Python library documentation for more advanced usage.
Configuration
See the APM Python library documentation for all the available configuration options.
Log Prompt & Completion Sampling
To enable log prompt and completion sampling, set the DD_LANGCHAIN_LOGS_ENABLED=1
environment variable. By default, 10% of traced requests will emit logs containing the prompts and completions.
To adjust the log sample rate, see the APM library documentation.
Note: Logs submission requires DD_API_KEY
to be specified when running ddtrace-run
.
Validation
Validate that the APM Python library can communicate with your Agent using:
You should see the following output:
Debug Logging
Pass the --debug
flag to ddtrace-run
to enable debug logging.
This displays any errors sending data:
ERROR:ddtrace.internal.writer.writer:failed to send, dropping 1 traces to intake at http://localhost:8126/v0.5/traces after 3 retries ([Errno 61] Connection refused)
WARNING:ddtrace.vendor.dogstatsd:Error submitting packet: [Errno 61] Connection refused, dropping the packet and closing the socket
DEBUG:ddtrace.contrib._trace_utils_llm.py:sent 2 logs to 'http-intake.logs.datadoghq.com'
Data Collected
Metrics
langchain.request.duration (gauge) | Request duration distribution. Shown as nanosecond |
langchain.request.error (count) | Number of errors. Shown as error |
langchain.tokens.completion (gauge) | Number of tokens used in the completion of a response. Shown as token |
langchain.tokens.prompt (gauge) | Number of tokens used in the prompt of a request. Shown as token |
langchain.tokens.total (gauge) | Total number of tokens used in a request and response. Shown as token |
langchain.tokens.total_cost (count) | Estimated cost in USD based on token usage. Shown as dollar |
Events
The LangChain integration does not include any events.
Service Checks
The LangChain integration does not include any service checks.
Troubleshooting
Need help? Contact Datadog support.