Skip to main content
This tutorial shows how to set up Keywords AI tracing with BAML to automatically track LLM usage from your BAML client and view detailed spans in the Keywords AI platform.
Import the instrumentation module before any other module imports baml_client.b to ensure all functions are decorated.

​
Setup

​
1) Environment Variables

Create a .env or environment variables in your runtime and set:
.env
KEYWORDSAI_API_KEY=your-keywordsai-api-key
KEYWORDSAI_BASE_URL=https://api.keywordsai.co

​
2) Create the Keywords AI BAML instrumentation module

Create a module (for example keywordsai_baml_tracing.py) that decorates all callable BAML client functions. Import this module first in your app so every BAML call is traced.
keywordsai_baml_tracing.py
from collections.abc import Callable
from functools import wraps

from baml_client import b
from baml_py import Collector
from keywordsai_tracing.decorators import task
from opentelemetry.semconv_ai import LLMRequestTypeValues, SpanAttributes
from servicelib.errors import logger
from servicelib.globals import kai_client

@task(name="baml_usage")
async def update_span_with_baml_usage(baml_function: Callable, *args, **kwargs):
    baml_collector = Collector(name="baml_usage_collector")
    kwargs.setdefault("baml_options", {})
    kwargs["baml_options"]["collector"] = baml_collector

    result = await baml_function(*args, **kwargs)

    input_tokens = baml_collector.last.usage.input_tokens if baml_collector.last and baml_collector.last.usage.input_tokens else 0
    output_tokens = baml_collector.last.usage.output_tokens if baml_collector.last and baml_collector.last.usage.output_tokens else 0

    baml_model_name = (
        baml_collector.last.selected_call.client_name
        if baml_collector.last and baml_collector.last.selected_call
        else None
    )

    attributes = {
        SpanAttributes.LLM_USAGE_PROMPT_TOKENS: input_tokens,
        SpanAttributes.LLM_USAGE_COMPLETION_TOKENS: output_tokens,
        SpanAttributes.LLM_USAGE_TOTAL_TOKENS: input_tokens + output_tokens,
        SpanAttributes.TRACELOOP_SPAN_KIND: LLMRequestTypeValues.CHAT.value,
    }
    if baml_model_name:
        attributes[SpanAttributes.LLM_REQUEST_MODEL] = baml_model_name
        attributes[SpanAttributes.LLM_RESPONSE_MODEL] = baml_model_name

    kai_client.update_current_span(
        keywordsai_params={"metadata": {"example": "baml"}},
        attributes=attributes,
        name="baml_usage.chat",
    )
    return result


def keywordsai_baml_trace(func: Callable):
    @wraps(func)
    async def wrapper(*args, **kwargs):
        return await update_span_with_baml_usage(func, *args, **kwargs)
    return wrapper


def initialize_keywordsai_baml_tracing() -> tuple[int, int]:
    """Apply keywordsai_baml_trace decorator to all callable BAML functions."""
    decorated_count = 0
    found_count = 0
    logger.info("[BAML Tracing] Decorating BAML client with tracing...")
    for attr_name in dir(b):
        if not attr_name.startswith("_"):
            attr = getattr(b, attr_name)
            if callable(attr):
                found_count += 1
                setattr(b, attr_name, keywordsai_baml_trace(attr))
                decorated_count += 1
    logger.info(f"[BAML Tracing] Decorated {decorated_count} / {found_count} BAML functions")
    return decorated_count, found_count

# Initialize immediately when this module is imported
_decorated_count, _found_count = initialize_keywordsai_baml_tracing()

# Export the decorated BAML client for other modules to use
__all__ = ["b"]

​
3) Use the decorated BAML client

Import the decorated b from your instrumentation module and use it like your normal BAML client:
app.py
from keywordsai_baml_tracing import b

async def main():
    # Call any BAML client function; usage will be traced automatically
    result = await b.some_function(arg1="value")
    print(result)

​
Observability

Once configured, you can view your traces and analytics in the Keywords AI platform under Signals → Traces.
⌘I