llmtrace-lite

Zero-config decorator for tracing LLM calls. Print-debugging for language model functions, nothing more.

What this is

A tiny Python decorator that logs basic metadata about LLM function calls: timing, model name, prompt/output sizes, and success/failure status. No dependencies, no configuration files, no dashboards. Just readable logs.

Installation

pip install llmtrace-lite

Usage

from llmtrace_lite import trace

@trace
def call_llm(prompt, model="gpt-4o"):
    # Your LLM call here
    response = some_llm_client.complete(prompt, model=model)
    return response

# Use it normally
result = call_llm("Explain quantum computing", model="gpt-4o")

Output:

[llmtrace] call_llm
  model: gpt-4o
  latency_ms: 842
  prompt_chars: 312
  output_chars: 1187
  status: success

File logging

Set LLMTRACE_FILE to append JSON lines instead:

export LLMTRACE_FILE=traces.jsonl
python your_script.py

Each line is a JSON object with trace metadata.

What this captures

• Function name
• Start/end timestamps
• Latency (milliseconds)
• Status (success or error)
• Exception details (if error)
• Model name (if passed as model kwarg)
• Prompt size (characters, if arg named prompt)
• Output size (characters, if return is string)
• Retry count (if function has retries attribute)

What this is NOT

This is not:

• An observability platform
• A provider-specific SDK integration
• An async framework
• A monitoring dashboard
• A production tracing system
• Extensible middleware

If you need those things, use LangSmith, Phoenix, or OpenTelemetry.

This is for local debugging and quick sanity checks. That's it.

Requirements

• Python 3.9+
• Standard library only
• No external dependencies

License

MIT