Better handling of score -1

[wdp-2010]

Describe the feature or potential improvement

I use score -1 to indicate tool calls, because they don't belong in the scorebook and therefore must be ignored
BUT -1 is being used in the avg and in some graphs

My question is if this is able to be implemented or is there an existing way of giving smth NO score and therefore ignore it?

Additional information

No response

[marliessophie]

Hey, could you please share more context on your workflow? What's your goal? To "count" tool calls? How many tools are available and how many were called? Or something different?

[wdp-2010]

Dear That’s for reaching outMy goal is to exclude tool calls from llm as a judge For example my judge is evaluating how close the ai is sticking to the system promp for user interaction, tools not I learned it to set the score to -1 when it sees it’s a tool call, this works but then the score spread graph doesn’t span from 0 to 1 but -1 to 1I would like to see a way to eliminate tool calls from the scores and avg and sutch metricsKind regards On 17 Dec 2025, at 15:01, marliessophie ***@***.***> wrote: Hey, could you please share more context on your workflow? What's your goal? To "count" tool calls? How many tools are available and how many were called? Or something different? —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you authored the thread.Message ID: ***@***.***>

[marliessophie]

Sorry, I cannot follow this, could you rephrase please?

[wdp-2010]

Rephrased by ai fact checked by me:

Thanks for reaching out.

My goal is to exclude tool calls from an LLM acting as a judge. For example, my judge evaluates how closely the AI follows the system prompt for user interaction, not its use of tools.

At the moment, I assign a score of -1 when a tool call is detected. While this works, it causes the score distribution graph to range from -1 to 1 instead of the intended 0 to 1.

Ideally, there would be a way to fully exclude tool calls from scoring, as well as from averages and related metrics. One possible approach would be for the judge to output a dedicated sentinel value (such as a specific symbol or a reserved value like -1) that downstream aggregation logic explicitly treats as “excluded” rather than as part of the numerical score range.

Kind regards

[marliessophie]

Would targeting an observation rather than a trace solve this?

[wdp-2010]

Can you define what an observation is?

[marliessophie]

Please refer to our core data model docs for context: https://langfuse.com/docs/observability/data-model