Problem scoping and metrics

We want to design a system for training SLMs for almost realtime evaluation of traces from GenAI applications which generate multimodal traces (sample use cases in Appendix). These evaluations serve as guardrail metrics which can monitor application performance and even trigger failsafe mechanisms

Guardrail metrics

We focus on extending common guardrail metrics to work in a multimodal setting. (Definitions in Appendix) The list is not exhaustive, but is representative of the kind of metrics we need to compute over the multimodal traces. Broadly, there are two different kind of metrics

1. (Input - Output) Correspondence metrics
   1. These require measuring some correspondence between inputs and outputs.
   2. Eg: MM Context adherence, Audio Chunk Attribution, Audio Chunk Utilization, MM Completeness, MM Generation Instruction adherence
2. (Input/Output) classification metrics
   1. These are computed by independently classifying a multimodal input or output into a set of relevant classes.
   2. Eg: Audio Tone Classification, PII-Image, PII-Audio

Scope exclusions

1. Video and code comprehension and generation applications
2. We only include guardrails which require decision making over multimodal inputs/outputs
3. Additionally, we only focus on guardrails for which it makes sense to use an SLM (some guardrails may be implemented in code or require LLM calls)

Modelling approaches

Input to evaluator

As discussed above, the metrics of interest can be viewed as either defined over an (image,text) pair or (audio,text) pair (correspondence metrics) or they can jus be defined over an image/audio.

In order to construct these inputs we need to pre-process traces and extract these. We will discuss this later in the doc and assume we we have the pre-processing done.

Baseline approaches

Encoder + learned similarity baseline (baseline1)

A simple baseline approach is training metric specific similarity models on top of frozen embeddings. For correspondence metrics, the simplest method would be to compute fixed dimension embedding (CLIP or SigCLIP for images, CLAP for audio and Qwen3 for text embedding) and concatenate them and pass it to a MLP with classification/regression objective.