We experienced the need for this when we were building a copilot for bankers. Our RAG pipeline blew up in complexity as we added components: a query classifier (to triage user intent), multiple retrievers (to grab information from different sources), filtering LLM (to rerank / compress context), a calculator agent (to call financial functions) and finally the synthesizer LLM that gives the answer. Ensuring reliability became more difficult with each of these we added.

When a bad response was detected by our answer evaluator, we had to backtrack multiple steps to understand which component(s) made a mistake. But this quickly became unscalable beyond a few samples.

I did my Ph.D. in fault detection for autonomous vehicles, and I see a strong parallel between the complexity of autonomous driving software and today's LLM pipelines. In self-driving systems, sensors, perception, prediction, planning, and control modules are all chained together. To ensure system-level safety, we use granular metrics to measure the performance of each module individually. When the vehicle makes an unexpected decision, we use these metrics to pinpoint the problem to a specific component. Only then we can make targeted improvements, systematically.

Based on this thinking, we developed the first version of continuous-eval for ourselves. Since then we’ve made it more flexible to fit various types of GenAI pipelines. Continuous-eval allows you to describe (programmatically) your pipeline and modules, and select metrics for each module. We developed 30+ metrics to cover retrieval, text generation, code generation, classification, agent tool use, etc. We now have a number of companies using us to test complex pipelines like finance copilots, enterprise search, coding agents, etc.

As an example, one customer was trying to understand why their RAG system did poorly on trend analysis queries. Through continuous-eval, they realized that the “retriever” component was retrieving 80%+ of all relevant chunks, but the “reranker” component, that filters out “irrelevant” context, was dropping that to below 50%. This enabled them to fix the problem, in their case by skipping the reranker for certain queries.

We’ve also built ensemble metrics that do a surprisingly good job of predicting user feedback. Users often rate LLM-generated answers by giving a thumbs up/down about how good the answer was. We train our custom metrics on this user data, and then use those metrics to generate thumbs up/down ratings on future LLM answers. The results turn out to be 90% aligned with what the users say. This gives developers a feedback loop from production data to offline testing and development. Some customers have found this to be our most unique advantage.

Lastly, to make the most out of evaluation, you should use a diverse dataset—ideally with ground truth labels for comprehensive and consistent assessment. Because ground truth labels are costly and time-consuming to curate manually, we also have a synthetic data generation pipeline that allows you to get started quickly. Try it here (https://www.relari.ai/#synthetic_data_demo)

What’s been your experience testing and iterating LLM apps? Please let us know your thoughts and feedback on our approaches (modular framework, leveraging user feedback, testing with synthetic data).