Code Integration

Import the required modules, SDK components, and constants.

import os
from logspend_sdk.core import LogBuilder, LogSpendLogger
from logspend_sdk.constants import Provider

Retrieve the API key from the environment variable.

api_key = os.environ.get('LOGSPEND_API_KEY')
if not api_key:
    raise ValueError("Missing LogSpend API Key. Ensure the LOGSPEND_API_KEY environment variable is set.")

Initialize the LogSpend Logger.

logger = LogSpendLogger(api_key=api_key, project_id="<YOUR_LOGSPEND_PROJECT_ID>")

Prepare your input data and instantiate the LogBuilder before calling the LLM API. Input data is identical to what the OpenAI API expects.

input_data = {
    "provider": Provider.OPENAI.value,
    "model": "gpt-3.5-turbo-instruct",
    "prompt": "Say this is a test", # Not needed for OpenAI integrations as everything is contained in the messages array below
    "messages": [
        {
          "role": "system",
          "content": "You are a helpful customer support assistant"
        },
        {
          "role": "user",
          "content": "Hello, can I get help with my booking"
        },
        {
          "role": "assistant",
          "content": "",
          "function_call": {}
        }
      ],
    "functions": [{}],
    "max_tokens": 7,
    "temperature": 0
}

builder = LogBuilder(input_data)

Set the identity on the LogBuilder instance. This helps to identify the user initiating the interactions. This must contain a session identifier session_id to encapsulate all interactions belonging to the same user session. This could be a chat ID, conversation ID, document ID or any relevant entity ID for your use case. Additionally, you can also include a user_id to represent the user (this could be an internal identifier, an email address, device ID, cookie identifier or any relevant unique ID to identify the user or visitor). Same applies to the optional ip_address field useful for enriching logs with user geographic information like city, country, and region.

identity_data = {
    "session_id": "session-123",
    "user_id": "123455",'
    "ip_address": "0.0.0.0",
}

builder.set_identity(identity_data)

Set any relevant custom properties on the LogBuilder instance. These custom properties can be used to filter on the data or create custom metrics and graphs on the LogSpend dashboard. Examples of custom properties are task_name, customer_id, etc.

custom_properties_data = {
    "task_name": "chatbot-qa",
    "customer_id": "chatbot-qa",
}

builder.set_custom_properties(custom_properties_data)

Make a call to OpenAI (or whichever provider you're using) to generate the output_data.

def call_openai(input_data):
    # Placeholder for actual OpenAI call
    # The fields below map well to the typical output of an OpenAI call with the exception of the http status code and optional error message
    return {
        "id": "chatcmpl-123",
        "choices": [{
                    "index": 0,
                    "message": {
                      "role": "assistant",
                      "content": "\n\nHello there, I'm happy to assist you today. What is your preferred destination?",
                    },
                    "finish_reason": "stop"
                 }]
                "usage": {
                    "prompt_tokens": 9,
                    "completion_tokens": 12,
                    "total_tokens": 21
                },
                "http_status_code": 200,
                "http_error_message": "",
    }

output_data = call_openai(input_data)

Set the output on the LogBuilder instance only after the API call to the LLM. Setting this in this order enables us to reliably estimate the latency. Output data is almost identical to what the OpenAI API returns.

builder.set_output(output_data)

Send the log to LogSpend.

logger.send(builder.build())