# Max Raw Score: normalizing retrieval before re-ranking > Max Raw Score caps raw retrieval scores before re-ranking, keeping the scoring pipeline calibrated and outliers from distorting results. **Category:** Retrieval Grounding **Author:** NeuralSeek Team · **Published:** June 9, 2026 **Canonical:** https://neuralseek.ai/ai-grounded/max-raw-score **Section index:** https://neuralseek.ai/ai-grounded Raw retrieval scores are noisy by nature. Most matches cluster in a sensible range, but every now and then an anomalous document scores wildly high — not because it's the best answer, but because of a quirk in how the query happened to overlap with the text. Left unchecked, that single outlier can dominate the ranking for entirely the wrong reasons. Max Raw Score caps those raw values before the re-ranking stage, so the pipeline stays calibrated and one freak score can't hijack the entire result set. ## What it actually does Before documents move into semantic re-ranking, their raw scores are clamped to a configurable maximum. This normalizes the input that re-ranking receives, preventing extreme values from flattening the differences between everything else. With the outliers brought back into range, the re-ranker can do its job — judging genuine semantic relevance — instead of being anchored by a number that was never meaningful to begin with. ## Why business teams care The payoff is stability, which is itself a form of trust. Without a cap, ranking can swing unpredictably whenever an anomalous match appears, so two nearly identical questions return noticeably different sources and the assistant feels arbitrary. Capping raw scores makes the system's behavior steadier and easier to rely on: the same kind of question reliably surfaces the same kind of source, which is exactly what users and auditors expect from a governed system. ## How to tune it in practice Most teams leave this near its default, because it is a calibration control rather than a behavioral one. The time to revisit it is when you notice ranking instability — similar questions returning surprisingly different sources, or one document repeatedly dominating where it shouldn't. Lowering the cap compresses the score range and tames aggressive outliers; raising it preserves more of the raw spread when your retriever's scores are already well-behaved. ## Common failure modes it prevents The core failure it guards against is 'outlier capture,' where a single anomalously high score crowds out genuinely strong-but-normal sources and the answer ends up grounded in the wrong place. A secondary benefit is protecting the re-ranker itself: feeding it wildly uncalibrated inputs makes its relevance judgments less reliable, so clamping the extremes upstream improves everything that depends on re-ranking downstream. ## Where it fits in the stack Max Raw Score sits between raw retrieval and semantic re-ranking — a quiet bridge that hands the re-ranker clean, bounded inputs. Because re-ranking, coverage weighting, and the confidence gates all build on this normalized foundation, a well-set cap quietly raises the reliability of the entire grounding stack. Most users never see this control directly, and that's the point: it works best when no one has to think about it. ## A pipeline-level safeguard Of all the guardrails, this is the most behind-the-scenes. It changes no visible behavior on a good day; it simply prevents the bad days where a single strange score would have made the assistant act erratically. That's what calibration controls do — they buy consistency, and consistency is what lets every more visible guardrail be trusted. > Consistency is a feature. An assistant that ranks the same question differently each time erodes trust faster than an occasional miss. ## The takeaway Max Raw Score keeps the scoring pipeline calibrated by capping outliers before re-ranking — a quiet, low-level safeguard that makes every grounded answer steadier and every downstream guardrail more dependable. --- From NeuralSeek's AI Grounded — practical, web-verified guidance on building governed, grounded enterprise AI. NeuralSeek is the model-agnostic, governed AI platform you own: any LLM (swap with no rebuild), your data in your own tenant (cloud or on-prem), 118 guardrails enforced before any action, one container that runs anywhere.