Most teams reach for the model picker first and the prompt last. That is backwards. On a grounded knowledge agent, the system prompt and the knowledge base account for the large majority of answer quality — the model is real, but it is the third lever, not the first.
1. Where the controls are
Open a bot and click Tune Bot in the sidebar, then the Behavior view. Three panels, plus a live preview of the bot beside them so you can test without leaving the page.
AI Model
Which model answers.
Instruction Prompt
Who the bot is and how it behaves.
Creativity
How far it strays from the retrieved text.
The Appearance view on the same page covers everything visual — that is covered in getting started.
2. Choosing a model
Chat.co offers models from OpenAI, Anthropic, Google and xAI directly, plus a wide set of models hosted on AWS Bedrock — including open-weight families. The picker groups them by provider.
How to choose, in practice
| Situation | What to reach for |
|---|---|
| Ordinary support and FAQ answering | A current mid-tier model. Fast, cheap per message, and on grounded questions the answer is mostly determined by retrieval anyway. |
| Long documents, synthesis, nuanced judgement | A frontier-tier model. Worth the cost when the work is genuinely hard rather than merely long. |
| Very high volume, latency-sensitive | The smallest tier in a current family. Noticeably faster, and often indistinguishable on simple retrieval questions. |
| Strict data-handling requirements | Filter to no-training models — see the next section. |
3. Model privacy posture
Every model in the picker carries a badge describing what happens to the prompts and responses that pass through it. This is a real, checkable property, not marketing language.
| Posture | What it means |
|---|---|
| No training | The provider is contractually excluded from training on your prompts and responses. Every Bedrock-hosted model carries this, because AWS Bedrock’s terms exclude customer content from training across all hosted foundation models. |
| No training, shared | Not trained on, but prompts and responses are shared with the model provider and retained for a limited window for abuse detection and possible human review. |
| Unknown | Chat.co has not made a classification for this route. Direct-API access to OpenAI, Anthropic, Google and xAI sits here — not because training is happening, but because the classification is deliberately withheld until the contractual language has been reviewed. |
A no-training only filter on the model picker hides everything that is not in the strictest category — including the “shared” variant, deliberately, since the filter promises the strongest posture available rather than an approximation of it.
The broader data-handling picture is in security & compliance.
4. The instruction prompt
The highest-leverage field in the product. It defines who the bot is, what it does when it does not know, what it refuses, and how it sounds. You can save named prompts and switch between them, and there is an Enhance action that rewrites a rough draft into something more specific.
What a good prompt covers
- 1
Identity and scope
Who the bot represents and what it is for. “You are the support assistant for Acme's billing platform” is a boundary, not just a label.
- 2
What to do when the answer isn't there
The most valuable instruction you can write. Say explicitly: admit it, do not improvise, offer a human. Without this the model will fill the gap, because filling gaps is what it is good at.
- 3
Hard refusals
Anything it must never do — quote prices it cannot verify, give legal or medical advice, speculate about a competitor, discuss another customer's account.
- 4
Tone and length
Be concrete. “Two or three sentences unless asked for detail” is actionable; “be friendly and professional” is not.
- 5
Escalation triggers
When to offer handoff: account questions, repeated failure, any sign of frustration, any direct request for a person.
You are the support assistant for Acme's billing platform. You answer using only the knowledge base provided to you. When you cannot answer from the knowledge base, say so plainly and offer to connect the person with our team. Never guess at pricing, contract terms, or the state of someone's account — you cannot see account data. Keep answers to two or three sentences unless asked for more detail. Cite your sources. Offer a human whenever: the question is about a specific account, you have failed to answer twice, the person sounds frustrated, or they ask for one. Never argue about a request to speak to a person.
5. Creativity
A slider from 0 to 100, mapping to the model's temperature. It controls how far the model is willing to move away from the most probable continuation — in practice, how far it strays from the text it retrieved.
| Range | Behaviour | Fits |
|---|---|---|
| 0–20 | Sticks closely to retrieved text; repeatable answers | Support, documentation, compliance, anything factual |
| 20–50 | Some rephrasing and synthesis, still well anchored | General assistants, mixed factual and conversational use |
| 50+ | Varied, more inventive, less predictable | Brainstorming and ideation — rarely what a knowledge agent wants |
6. Letting visitors choose
Under Appearance → Chat Features you can expose a set of models to visitors, so they pick which one answers. Useful for internal tools and technical audiences who have their own preferences.
For a customer-facing support bot, usually not: it presents a choice most visitors have no basis to make, and it makes your answers less predictable across sessions in a way that is hard to debug. Pick a model, and own the result.
7. How to tune without guessing
The failure mode here is changing three things, seeing an improvement, and never knowing which one caused it.
- 1
Write down ten questions and their right answers
Real ones, in the words your customers use. This list is the only thing standing between tuning and superstition, and it takes twenty minutes to build.
- 2
Run them and record what you get
Use the live preview beside the controls, or Test Bot. Note which answers are wrong and, more importantly, how they are wrong.
- 3
Diagnose before you change anything
Wrong facts or missing citations mean the knowledge base is the problem. Right facts in the wrong shape — too long, too formal, hedging when it should decline — mean the prompt is. Inconsistent answers across runs mean creativity is too high. Only when none of these fits should you reach for the model.
- 4
Change one thing, re-run the same ten
One variable at a time. It is slower for one iteration and much faster over five, because you actually learn what worked.