Trace DeepSeek applications

​

Installation

pip install openai langsmith

​

Setup

export LANGSMITH_API_KEY="your-langsmith-api-key"
export LANGSMITH_TRACING="true"
export LANGSMITH_PROJECT="deepseek-integration"
export DEEPSEEK_API_KEY="your-deepseek-api-key"

• Ensure you have a DeepSeek API key from your DeepSeek account.
• Set LANGSMITH_TRACING=true and provide your LangSmith API key (LANGSMITH_API_KEY) activates automatic logging of traces.
• Specify a LANGSMITH_PROJECT name to organize traces by project; if not set, traces go to the default project (named “default”).
• The LANGSMITH_TRACING flag must be true for any traces to be recorded.

​

Configure tracing

1. Instrument the DeepSeek API call with LangSmith. In your script, create an OpenAI client configured to use DeepSeek’s API endpoint and wrap a call in a traced function:
   Python

   TypeScript

   Copy

   import os
   from openai import OpenAI
   from langsmith import traceable
   
   # Create a client pointing to DeepSeek
   client = OpenAI(
       api_key=os.environ["DEEPSEEK_API_KEY"],
       base_url="https://api.deepseek.com/v1"
   )
   
   @traceable(
       run_type="llm",
       name="DeepSeek Chat Completion",
       metadata={"ls_provider": "deepseek", "ls_model_name": "deepseek-chat"},
   )
   def call_deepseek(messages: list[dict]):
       response = client.chat.completions.create(
           model="deepseek-chat",
           messages=messages
       )
       return response.choices[0].message
   
   if __name__ == "__main__":
       messages = [
           {"role": "system", "content": "You are a helpful assistant that translates English to French."},
           {"role": "user", "content": "I love programming."}
       ]
       result = call_deepseek(messages=messages)
       print("Model reply:", result.content)
   
   In this example, you use the OpenAI SDK to interact with DeepSeek’s API. The OpenAI client is configured with base_url="https://api.deepseek.com/v1" to route requests to DeepSeek’s endpoint while maintaining OpenAI-compatible syntax. The @traceable decorator (Python) or traceable function (TypeScript) wraps your function so that each invocation is logged as a trace run of type "llm". The metadata parameter tags the trace with:
   • ls_provider: Identifies the provider (DeepSeek) for filtering traces.
   • ls_model_name: Specifies the model used for cost tracking and analytics.
   The function returns the full message object (response.choices[0].message), which includes the response content along with metadata like the role and any additional fields. LangSmith automatically captures:
   • Input messages sent to the model.
   • The model’s complete response (content, role, etc.).
   • Model name and token usage statistics.
   • Execution timing and any errors.
2. Execute your script to generate a trace:
   Python

   JavaScript

   Copy

   python deepseek_trace.py
   
   The function call will reach out to DeepSeek’s API, and because of the @traceable/traceable wrapper, LangSmith will log this call’s inputs and outputs as a new trace. You’ll find the model’s response printed to the console, and a corresponding run appear in the LangSmith UI.

​

View traces in LangSmith

1. Open the LangSmith UI and log in to your account.
2. Select the project you used for this integration (for example, the name set in LANGSMITH_PROJECT, or “default” if you didn’t set one).
3. Find the trace corresponding to your DeepSeek API call. It will be identified by the function name (DeepSeek Chat Completion).
4. Click on the trace to open it. You’ll be able to inspect the model input and output, including the prompt messages you sent and the response from DeepSeek, as well as timing information (latency) and token usage.

​

Cost tracking