Analyze and Debug LLM Requests

import os
import uuid
import traceback
import time

from openai import OpenAI
from autoblocks.tracer import AutoblocksTracer

openai_client = OpenAI()

tracer = AutoblocksTracer(
  # All events sent below will have this trace ID
  trace_id=str(uuid.uuid4()),
  # All events sent below will include this property
  # alongside any other properties set in the send_event call
  properties=dict(
    provider="openai",
  ),
)

params = dict(
  model="gpt-3.5-turbo",
  messages=[
    {
      "role": "system",
      "content": "You are a helpful assistant. You answer questions about a software product named Acme.",
    },
    {"role": "user", "content": "How do I sign up?"},
  ],
  temperature=0.7,
  top_p=1,
  frequency_penalty=0,
  presence_penalty=0,
  n=1,
)

# Use a span ID to group together the request + response events
span_id = str(uuid.uuid4())

tracer.send_event(
  "ai.request",
  span_id=span_id,
  properties=params,
)

try:
  start_time = time.time()
  completion = openai_client.chat.completions.create(**params)
  tracer.send_event(
    "ai.response",
    span_id=span_id,
    properties=dict(
      # OpenAI returns pydantic models, so they need
      # to be serialized via model_dump.
      response=completion.model_dump(),
      latency_ms=(time.time() - start_time) * 1000,
    ),
  )
except Exception as error:
  tracer.send_event(
    "ai.error",
    span_id=span_id,
    properties=dict(
      error=dict(
        type=type(error).__name__,
        message=str(error),
        stacktrace=traceback.format_exc(),
      ),
    ),
  )
  raise

# Simulate user feedback
tracer.send_event(
  "user.feedback",
  properties=dict(
    feedback="good",
  ),
)