# Strict, Balanced & Open Grounding Modes | onmsg

> Compare onmsg

URL: https://onmsg.app/guide/strict-balanced-and-open-grounding-modes-compared/
Last-Modified: 2026-09-08

Decision guide

# Strict, Balanced and Open Grounding Modes

Compare onmsg's grounding modes: Strict (refuse when no relevant content), Balanced (prefer your knowledge), and Open (general conversation), plus when each fits.

Published September 8, 2026 · 5 min read

![Three labelled dials representing Strict, Balanced and Open grounding modes](/images/featured/modern-3d-illustration-of-three-labelled-dials-for.webp)

## One setting, three behaviours

Grounding mode decides what the 

Grounded AI Agent & Knowledge Base

[/features/grounded-ai-agent/ →](/features/grounded-ai-agent/)

 does when a visitor asks something. Specifically, it decides what happens in the gap between “your knowledge base has an answer” and “the model could produce one anyway”.

Three modes, and the difference between them is entirely about that gap.

![A comparison table of the three grounding modes with a when-to-use column](/images/content/comparison-table-of-the-three-grounding-modes-with.webp)

## Strict

Strict mode checks for relevant knowledge first. If it finds none, it can return your configured refusal message **without calling the answer model at all**. The model never gets the chance to improvise, because it is never asked.

That is a stronger guarantee than a prompt instruction, and it is the reason Strict is the right default for anything with a number or a promise in it. Prices, coverage areas, lead times, warranty terms, eligibility rules. The worst case becomes “I can’t confirm that, let me get someone”, which is an answer you can live with.

The cost is refusals on questions you could have answered if the knowledge had been there. Treat those as a to-do list rather than a fault: every refusal in your transcripts is a missing knowledge source, precisely identified.

## Balanced

Balanced mode prefers the knowledge you supplied and instructs the model not to invent business facts, while still letting it answer conversationally around your content. It does not gate as hard as Strict, so it will engage with a question that only partly matches your material.

This suits general support and explanatory questions where tone matters and the downside of a slightly loose answer is small. “How does your process usually work?” is a Balanced question. “What will it cost me?” is not.

## Open

Open mode allows a more general conversation without the retrieval gate. Use it where the chat is a companion rather than an authority, a general assistant on a content site, an exploratory conversation with no commercial edge.

For a small business website selling something, Open is rarely the right pick, and we would rather say that plainly than list it as an equal option.

## Side by side

|  | Strict | Balanced | Open |
| --- | --- | --- | --- |
| Checks knowledge first | Yes, gating | Yes, preferring | No gate |
| When nothing relevant is found | Configured refusal | Answers cautiously | Answers generally |
| Best for | Prices, promises, eligibility | General support, process questions | Conversational, low-stakes |
| Main risk | Refusing answerable questions | Looser answers on edge cases | Fabricated business facts |
| Main fix | Add the missing knowledge | Tighten retrieval strictness | Switch mode |

## Retrieval strictness sits on top

![A retrieval strictness slider being adjusted beside a configured refusal message](/images/content/modern-3d-illustration-of-a-retrieval-strictness-s.webp)

Whichever mode you pick, retrieval strictness is adjustable. It controls how closely a passage must match the question before it counts as relevant.

Tighten it when you see answers built from loosely related content, a passage about one service being used to answer about another. Loosen it when reasonable questions are being refused despite the knowledge existing, which usually means your sources are phrased differently from the way customers ask.

Two knobs, and they interact. Strict with very tight retrieval refuses a lot. Balanced with loose retrieval drifts. Most setups land on Strict with moderate retrieval, then adjust after reading a week of transcripts.

## Mixing modes across a site

You are not choosing one mode for the whole business. Agents are configured individually, so a flow can use a Strict agent on the pricing branch and a Balanced one on the general-questions branch. That is usually a better answer than compromising in the middle.

## Configure the refusal at the same time

A mode is only half the decision. The other half is what the visitor sees when the gate closes. Write a refusal message that says what you cannot confirm and offers the next step, then set the number of refusals that should trigger a handoff to a person.

## Diagnosing a mode that is set wrong

Four symptoms map cleanly onto four fixes.

**Refusing questions you clearly documented.** The knowledge exists but is not being retrieved. Either loosen retrieval strictness, or rewrite the source using the words customers actually use. The second is usually the better fix.

**Answering adjacent questions.** The agent responds about a related service rather than the one asked about. Retrieval is too loose, or two sources are too similar to distinguish. Tighten strictness first, then look at whether the sources need clearer boundaries.

**Confident answers about things you never wrote down.** You are in Balanced or Open where you needed Strict. Change the mode.

**Refusing everything.** Usually an empty or barely populated knowledge base rather than a mode problem. Add content before adjusting settings.

The general rule: symptoms about _what_ is retrieved are strictness problems, symptoms about _whether the model answers anyway_ are mode problems.

## Reviewing your choice after launch

Set the mode on day one, then revisit it once you have a fortnight of real conversations. Two questions are enough.

Would any answer you saw have embarrassed you if a customer acted on it? If yes, that branch needs Strict.

Were there refusals a person would have answered easily from material you already publish? If yes, the fix is content rather than configuration, add the source, keep the mode.

Most businesses tighten after the first review and then leave it alone. This is not a setting that needs continuous tuning once the knowledge base is honest.

What happens when the AI can’t answer

[/guide/what-happens-when-the-ai-cant-answer/ →](/guide/what-happens-when-the-ai-cant-answer/)

 covers that side in detail, and 

how to reduce AI chatbot hallucinations

[/guide/how-to-reduce-ai-chatbot-hallucinations/ →](/guide/how-to-reduce-ai-chatbot-hallucinations/)

 puts the whole set of controls in context.

## Learn more about Grounded AI Agent & Knowledge Base

An AI agent that answers from your own knowledge base, with grounding modes, configurable refusals, and handoff to a person.

Read the feature page

[/features/grounded-ai-agent/ →](/features/grounded-ai-agent/)

FAQ

## Questions people ask about this

### Which mode is safest?

Strict, because it checks for relevant knowledge and can return your configured refusal without calling the answer model when it finds none. A refusal you wrote is a predictable outcome; a generated guess is not.

### When would I use Open mode?

Where the chat is conversational rather than authoritative and there is no cost to a general answer. If prices, availability or promises are involved, Open is the wrong choice.

### Can I tune how strict retrieval is?

Yes. Retrieval strictness is adjustable independently of the mode, so you can decide how closely a passage must match before it counts as relevant. Tighten it if you see loosely related passages being used; loosen it if valid questions are being refused.

## Related guides

### How to Add Your Business Knowledge: Paste, Upload and URL Import

Supply knowledge by pasting text, uploading PDF/DOCX/TXT/Markdown/CSV, or importing one public web page. See processing status, reindex and edit sources.

Read guide

[How to Add Your Business Knowledge: Paste, Upload and URL Import →](/guide/how-to-add-your-business-knowledge/)

### How to Reduce AI Chatbot Hallucinations with Grounding Controls

Why chatbots invent facts, and how strict retrieval gating, refusal thresholds, handoff and source attribution help reduce it, controls that guide behaviour.

Read guide

[How to Reduce AI Chatbot Hallucinations with Grounding Controls →](/guide/how-to-reduce-ai-chatbot-hallucinations/)

### Website Data Privacy: What Visitors' Chat Data Means for GDPR and CCPA

What chat and enquiry data is collected, how retention settings differ by plan, and general GDPR/CCPA considerations for chat intake, guidance, not legal advice.

Read guide

[Website Data Privacy: What Visitors' Chat Data Means for GDPR and CCPA →](/guide/website-data-privacy-gdpr-and-ccpa/)

### What Does "Grounded AI" Mean for Website Chat?

Grounded AI answers from your own content instead of general knowledge. What grounding is, why it matters, and how it differs from a general chatbot.

Read guide

[What Does "Grounded AI" Mean for Website Chat? →](/guide/what-does-grounded-ai-mean-for-website-chat/)

## Want to try this on your own site?

No credit card required.

Start free

[https://chat.onmsg.app →](https://chat.onmsg.app)