Adding structure to traces
Default automatic tracing
If you initialize Laminar, LLM calls will be automatically traced. Typically, one LLM call creates one span.
Here, openai.chat
is a span. It is the only span in the trace.
Since these spans are created automatically, they are not grouped into traces. That is, you will see one span per trace.
For example, in this toy application, we make two LLM calls – first generates random nouns, and second generates a poem using those nouns.
Here, both entries in the table are traces. Each trace contains one span.
On the right-hand side, you can see the expanded view of the trace. It shows
the only span openai.chat
.
It will make sense to group these spans into a single trace, as they belong to the same user request.
Grouping spans into traces
The most common way to group spans into traces is to create one parent span before making LLM calls. This parent span will contain all the child spans, each may be an LLM call, or a piece of code that you want to trace.
The concept of such a parent span is sometimes called a “root span” or a “top-level span”.
The recommended way to create a parent span is to use the observe
decorator (Python)
or observe
function wrapper (JavaScript/TypeScript).
You can instrument specific functions by adding the @observe()
decorator.
This is especially helpful when you want to trace functions, or group
separate functions into a single trace.
We are now recording my_function
and the OpenAI call, which is nested inside it, in the same trace. Notice that the OpenAI span is a child of my_function
. Parent-child relationships are automatically detected and visualized with tree hierarchy.
You can nest as many spans as you want inside each other. By observing both the functions and the LLM/vector DB calls you can have better visualization of execution flow which is useful for debugging and better understanding of the application.
Grouping traces into sessions
Sometimes, you may want to group traces into sessions. Sessions are useful when you want to group traces that are related to a single user interaction or a single request.
For example, imagine an advanced conversational chatbot agent that at every turn in the conversation does several things, such as processing the user input, calling the database, and generating a response. In such case, we may want to represent each turn in the conversation as a trace, and the whole conversation as a session.
Example. Associate a trace with a session
Simply call the Laminar.set_session(session_id="session123")
within or outside any span context.
All the subsequent spans will be associated with the session.
Viewing sessions
Head to the traces page and select the “Sessions” tab. You will see all the sessions, and if you click on each, it will expand to show all the traces within the session.
Each trace within a session is the same as if it were a standalone trace.
Dynamically disabling tracing
Sometimes, you may want to dynamically disable tracing for a particular span. For example, some of your customers need more privacy than others, and you only want to collect metadata for some of them.
To achieve this, we offer a wrapper that can set tracing to one of the following modes:
ALL
– trace everything as normalMETA_ONLY
– do not trace inputs and outputsOFF
– do not trace anything within the wrapper
Next Steps
Learn more about observe
and its advanced alternatives in the section on
Manual instrumentation.