The Assistant
This section is about the part of Promptly that thinks. Your knowledge base and catalog are what the assistant knows; the widget is where visitors meet it. What's in between — which model runs, what it's told to be, how it plans an answer, and what it remembers — is here.
Most of it works out of the box. The pages worth reading on day one are Writing a system prompt and Model & behavior settings; the rest is there for when you want to understand an answer you didn't expect.
The shape of an answer
Every visitor message goes through roughly the same arc, and knowing it makes the rest of this section make sense:
- Classify — what kind of question is this? A greeting, a product search, a policy question, a request for a human? See Intent routing.
- Plan — for anything non-trivial, build a short sequence of steps: search this, filter that, sort, check the knowledge base. See Multi-step reasoning.
- Run — execute those steps against your own data, and any custom tools you've configured.
- Compose — hand everything gathered to the model, which writes one answer in the visitor's language, with product cards where they help.
The visitor sees only step 4. Steps 1–3 are invisible and usually take a couple of seconds.
Two rules hold across all of it: the assistant answers from your data and won't invent a product, price, or fact to fill a gap; and everything it touches is scoped to your workspace alone.
In this section
- Model & behavior settings — which model answers (it comes from your own API key), plus temperature, reply length, and how much conversation history travels with each message.
- Writing a system prompt — the standing instruction that gives your bot a name, a tone, and its rules. Includes the one phrasing mistake that quietly breaks working bots.
- Multi-step reasoning — how a compound question ("3-bedroom near the centre under €200k with a fireplace") becomes a plan, gets executed, and comes back as one answer.
- Intent routing — the five verdicts every message is sorted into, what each one skips, and why an uncertain classification deliberately takes the slower path.
- Custom tools & API actions — let the assistant call your own APIs mid-conversation to fetch live data (order status, stock) or trigger an action.
- Cross-session memory — recognise returning visitors and reference what they asked about last week, not just earlier in this chat.
- Languages & cost tracking — automatic replies in the visitor's language, and where the cost of all this actually lands (your provider's bill, not ours).
If the answers aren't good
Diagnose in this order — it goes cheapest-to-fix first, and most problems are solved before the end of the list:
- Is the content there at all? Most "the bot is dumb" reports are missing knowledge. See Adding knowledge.
- Can it be found? Content phrased nothing like the questions people ask won't be retrieved. See How retrieval works.
- Does the bot know your vocabulary? For product questions, the mapping from customer words to your field values lives in AI context & examples.
- Is the tone or scope wrong? Now it's a system prompt problem.
- Is it too creative or too terse? Now it's temperature and max tokens.
Notice that model choice isn't on the list. Switching models rarely fixes a content problem — and when nothing above works, Analytics will show you the questions that actually failed.