Chatbot AI settings

What it is

AI configuration is the smallest settings tab in the product and, for once, that is the feature rather than an omission. It has exactly three controls:

1. Response Length — Short Response or Long Response.
2. Show reference links — an on/off switch for source links under answers.
3. Clickable reply options — an on/off switch that turns a short list of choices at the end of a reply into buttons visitors can tap.

Everything else people expect to find here — the temperature, a system-prompt box, a confidence threshold — is owned by the platform, not by you. The AI model itself is not on this tab either: it has its own AI model card on the chatbot's Playground tab. The rest of this page explains what each of the three controls actually does, how to choose, and why the others are absent.

When you would use it

• Answers feel too long for a chat bubble, or too curt for a documentation portal.
• You want visitors to be able to click through to the page an answer came from.
• You want the opposite: prose-only replies with no URLs at all.

Where to find it

Open Dashboard → Chatbots → your chatbot → Settings → AI configuration (?tab=ai).

Control 1 — Response Length

Two options, shown as cards:

Option	The card says	What it actually does
Short Response	"Concise, direct answers that get straight to the point. Faster, to-the-point replies."	Instructs the model to produce a short, concise response, with guidance that favours brevity while still covering the question
Long Response	"Detailed answers with examples and extra context. In-depth answers with more context."	Instructs the model to produce a detailed, thorough response, with explicit guidance not to pad the answer with invented material

Short Response is the default, and it is the right default for most people.

This is not a hard character cap. It is a length policy the model follows, tuned so each mode keeps its own risk in check: Short is steered away from dropping a necessary caveat to save words, and Long is steered away from inventing content just to fill the space. A question with a two-word answer gets a two-word answer in either mode.

Which one to pick

Choose Short when:

• The chat appears as a bubble on your site, where a wall of text feels like a wall of text.
• Your questions are factual and closed: opening hours, return windows, sizes, prices.
• Visitors are on phones.

Choose Long when:

• The chatbot explains procedures, eligibility rules, or anything with conditions attached.
• It sits on a documentation or support portal where people expect depth.
• Answers in Short mode are technically correct but keep triggering a follow-up question.

A useful test: ask your five hardest questions in Test Chatbot under Short. If you find yourself asking "but what about…" after most replies, switch to Long. If you find yourself skimming, stay Short.

Does length cost more?

A conversational turn costs the same number of credits either way — credits are charged per answer at the chatbot's AI model rate (1X = one credit), not per word. Long Response does generate more text, so a chatbot set to Long works through its monthly allowance no faster in credits, but produces meaningfully more output per turn. Set the length your audience needs and no higher. See Message credits.

Control 2 — Show reference links

When this is on and the chatbot has found matching knowledge, replies can include links back to the source, so a visitor can verify the answer or read the full page.

When it is off, the chatbot is instructed to produce a links-free, prose-only reply — it will not include URLs or a resources block at all. This is not merely hiding a chip in the interface; the answer text itself changes.

The constraint nobody expects

Links can only be shown for knowledge crawled from a website. A crawled page has an address to link to. Pasted text and uploaded files do not — there is no URL to point at, so no link appears no matter how the toggle is set.

So if your knowledge is entirely PDFs and pasted text, turning this on changes nothing visible, and that is correct behaviour rather than a bug. If you want clickable sources, crawl the website that hosts the same material.

When to turn it on

On — public documentation, policy answers, anything where a visitor may reasonably want to check you, and any case where reading the full page is genuinely more useful than the summary.

Off — when your sources are internal documents you would rather not expose the addresses of, when the chat is embedded somewhere links would look out of place, or when your knowledge has no crawled pages anyway.

Independently of this setting, you can always inspect which sources fed an answer, from the transcript in Conversations and inside Test Chatbot. This toggle governs what visitors see. See Sources and citations.

Control 3 — Clickable reply options

Sometimes the chatbot cannot answer until the visitor narrows things down — "Which plan do you mean?" followed by a short list of plans. With this switch on (the default), that trailing list of choices is shown as a row of buttons under the reply instead of as plain bullets. The visitor taps one and the text of that button is sent as their next message, exactly as if they had typed it — same conversation, same credit, same answer path.

It applies wherever the chatbot talks to people: the embedded widget, the hosted chat page, and your own Test Chatbot. The chatbot is also instructed to offer choices this way — as a short, unnumbered bullet list at the end of the reply — so the buttons appear when they are useful and not for every list.

What visitors see

• Only the latest reply is interactive. Once the visitor answers (by tapping or typing), the buttons go away and that earlier reply shows its list as ordinary text, so the transcript stays complete.
• A list becomes buttons only when it looks like a set of choices: 2 to 6 short, unnumbered bullets at the very end of a reply, after a sentence that ends with a colon or a question mark. Numbered steps, long items, and lists with links stay as text.
• While the reply is still streaming in, the list is plain text; the buttons appear once the reply finishes.

When to turn it off

Turn it off if your chatbot's answers frequently end in short lists that are not choices — a recap of features, say — and the buttons read as an invitation to pick one. Off means plain lists everywhere for that chatbot.

This is a display setting: switching it does not change the answer text and, unlike the two controls above, does not clear cached answers. Saving it does update the widget on your site within a few seconds.

Where the AI model is chosen

The model is not on this tab. Open the chatbot's Playground tab and use the AI model card beside the preview:

• On the Free plan every chatbot answers with the default Agentency AI model, at 1X — one credit per reply.
• On a paid plan select Change AI model to pick a live model for visitors and, optionally, a separate testing model for your playground and Test Chat. Each model shows its credit multiplier, from 1X to 12X — see Message credits.

Whichever model you pick, its sampling recipe — including temperature — is chosen and maintained by Agentency. There is no slider, and posting a model or temperature with the chatbot's settings through the API does nothing.

That is on purpose, and it buys you two things:

• Answer quality stays consistent across every chatbot on the platform, so support advice actually applies to your case.
• You never have to reason about temperature. Getting it wrong is one of the easiest ways to make a grounded chatbot start inventing things.

Similarly, the retrieval quality gate — how confident a match must be before it is used — is fixed platform-wide rather than being a per-chatbot slider.

You do not need an OpenAI key or any other provider key. The API keys under Settings → Developer are for calling the Agentency API from your own code; they have nothing to do with swapping models. See Personal access tokens.

What you actually control, and where

Since this tab is short, here is the honest map of what shapes an answer:

To change…	Go to
How long the answer is	This tab — Response Length
Whether visitors see source links	This tab — Show reference links
Whether a trailing list of choices becomes tappable buttons	This tab — Clickable reply options
Which AI model answers, and its credit rate	Playground tab — the AI model card
Tone, rules, banned topics	Write instructions
What the chatbot knows at all	Knowledge overview
The greeting and the "I don't know" reply	Chatbot messages and language
Things the chatbot can do	What are Call actions

Instructions and knowledge are where the leverage is. If answers are wrong, this tab is not the fix.

Steps

1. Open the chatbot → Settings → AI configuration.
2. Select Short Response or Long Response.
3. Switch Show reference links on or off. If you turn it on and your knowledge is all uploads and pasted text, note the reminder that links need crawled pages.
4. Leave Clickable reply options on unless your replies often end in lists that are not choices.
5. Select Save changes.
6. Open Test Chatbot from the chatbot header and ask something your knowledge genuinely covers.
7. Check both things at once: the reply is the length you wanted, and sources appear (or do not) as you intended. Ask something deliberately vague, too, and see the choices come back as buttons.

What you will see

A card with two response-length options and two toggles. No model list, no sliders. Save and Discard appear once you change something.

Because Response Length and reference links change the answer text, saving them clears any cached answers for this chatbot — so your next test really is a fresh answer rather than a replay. Clickable reply options is a display setting and, like cosmetic changes elsewhere (colours, sounds, the welcome message), deliberately does not clear that cache.

Limits and plan notes

Available on every plan, including Free. You need permission to edit chatbots.

Turning reference links on does not create knowledge. An untrained chatbot, or one with no crawled pages, still has nothing to cite. Each turn you spend testing costs message credits like any other, at the testing model's rate.

Common problems

I do not see a model dropdown or a temperature slider.

Expected on this tab. The model lives on the AI model card on the Playground tab (choosing one needs a paid plan); temperature and sampling are platform-owned everywhere.

I enabled reference links and no links appear.

The most likely cause is that your knowledge has no crawled web pages — uploads and pasted text have no address to link to. Also confirm the question was actually answered from knowledge, and that the source finished training. See Sources and citations.

Answers are still long after I chose Short Response.

Length is a policy, not a hard cap, and a question with a genuinely complex answer stays long in either mode. If you want a strict style, add an instruction such as "answer in at most three sentences" — see Write instructions.

Answers got shorter but also worse.

Short mode is calibrated not to drop necessary detail, so if replies lost something important the cause is usually thin knowledge rather than the length setting. Switch back to Long, confirm the content is actually there, then improve the source.

Where do I paste my OpenAI key?

Nowhere. There is no place to bring your own model. The Developer settings keys are for the Agentency API.