# Document Score Range: the relevance band that keeps answers on-topic > Document Score Range sets the relevance band for what gets pulled from your knowledge base — so the model only ever sees sources worth answering from. **Category:** Retrieval Grounding **Author:** NeuralSeek Team · **Published:** June 9, 2026 **Canonical:** https://neuralseek.ai/ai-grounded/document-score-range **Section index:** https://neuralseek.ai/ai-grounded Every grounded answer begins with retrieval, and retrieval is only as good as the sources it allows through. Document Score Range is the dial that defines the relevance band — the minimum and maximum match scores a document must earn to even enter the conversation. Get it right and the model reasons over genuinely relevant material; get it wrong and you either starve the answer of evidence or flood it with noise that drowns out the truth. It is the very first decision in the grounding pipeline, and because everything downstream inherits from it, it is also the most consequential. ## What it actually does When a question arrives, the knowledge base returns candidate documents ranked by how well they match the query. Document Score Range defines the acceptable window for those scores. Anything below the floor is too loosely related to trust and gets discarded before it can mislead the model. Anything above an unusually high ceiling is frequently a near-duplicate, a boilerplate header, or a navigation fragment that matches the query mechanically but adds no real information. By bounding both ends, the band keeps the model focused squarely on the sources that carry useful, answer-bearing content. ## Why business teams care This control is the difference between an assistant that answers from the correct policy document and one that confidently cites something tangential. Tightening the band raises precision — fewer off-topic answers — at the occasional cost of the assistant saying 'I don't have that information.' Loosening the band raises coverage but invites drift, where the model stretches weak matches into confident-sounding claims. That trade-off is fundamentally a business judgment about how cautious the assistant should be, and Document Score Range expresses it as a single, explicit, auditable setting rather than a mystery buried in code. ## How to tune it in practice Start in the middle and watch two signals: how often the assistant declines to answer, and how often it answers from something only loosely relevant. If decline rates are high and the knowledge base genuinely contains the answers, lower the floor a notch. If you see off-topic citations creeping in, raise it. The ceiling matters most for knowledge bases full of repetitive or templated content — lowering it suppresses those mechanical high-score matches so real sources can surface. Change one bound at a time and re-test against a fixed set of representative questions so you can attribute every shift in behavior. ## Common failure modes it prevents Without a sensible band, two classic failures appear. The first is the 'confident tangent,' where a low-relevance document scrapes past with just enough keyword overlap and the model treats it as authoritative. The second is the 'boilerplate echo,' where headers, disclaimers, or repeated footer text score artificially high and crowd out the substantive passages. A well-set Document Score Range neutralizes both, so the model spends its attention on material that can actually answer the question. ## Where it fits in the stack Document Score Range is the entry gate. Everything that follows — re-ranking, coverage weighting, semantic thresholds, confidence gates — operates on the documents this band lets through. That makes it a force multiplier: a clean, well-bounded candidate set makes every downstream hallucination control sharper and cheaper, while a sloppy one forces those controls to work harder to compensate for noise that should never have been admitted in the first place. ## Tuned per use case A legal or compliance assistant runs a tight band where precision is everything; a broad internal helpdesk can afford a wider one because the cost of an imperfect-but-helpful answer is low. Because this is a per-tenant setting, the same platform can serve a risk-averse banking workflow and a forgiving employee FAQ simultaneously, with no code changes and no separate deployment — just two different positions on the same dial. > Retrieval quality is destiny. If the wrong documents reach the model, no amount of clever prompting can save the answer. ## The takeaway Document Score Range is the first and most consequential gate in grounding: a tunable relevance band that decides what the model is even allowed to reason from — the clean foundation every downstream guardrail depends on. --- 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.