Thank you! I really appreciate the thoughtful feedback. I completely agree. A self-healing pipeline is only as good as the signals it uses to make retry decisions. In the current version, I'm using retrieval confidence, answer quality, and hallucination detection as the decision inputs, but they're still largely heuristic-driven. One of the upcoming milestones is to introduce a proper benchmarking and observability layer so I can measure retrieval quality, grounding, retry effectiveness, and overall healing success across representative datasets instead of relying solely on thresholds. The goal is to make every retry decision explainable and backed by evaluation metrics. You're also absolutely right about chunking and metadata. I've learned that retrieval quality is influenced far more by document preprocessing than by simply swapping embedding models. Thanks again for taking the time to share such detailed feedback!