Re:PREZIO
Getting started

Creating agents (basics)

In this section you create a new AI agent and define its core behaviour: what it's called, which AI model it uses, how creative or how detailed its responses are, which language it speaks and what role it actually takes on. You open it via "New agent" in the agent overview. Everything you configure here can be changed again at any time later.

Creation runs in two steps (you'll see a progress bar at the top, the "stepper"):

  1. Phone number — first you choose a dedicated phone number (buy a new one or take over an existing one from your account). This number powers Voice and WhatsApp for this agent. Phone setup is covered in detail in its own chapter; for the basics it's enough to select a number and click "Next".
  2. Agent details — here you choose a template and then fill out the actual configuration form. That's exactly what this chapter is about.

Step 1: Choose a template ("Choose template")

Before the form appears, Prezio asks: "What is this agent for?" The chosen template automatically sets up suitable capabilities and a starter prompt — you can still adjust everything in the form afterwards.

The available options are:

  • Restaurant — answers menu questions from the knowledge base and takes orders (ordering capability pre-enabled).
  • Appointments — books, reschedules and cancels appointments and answers service questions (booking capability pre-enabled).
  • Sales / Leads — qualifies and captures prospects and answers questions from the knowledge base.
  • Support — answers questions strictly from the knowledge base and otherwise escalates to a human.
  • Custom ("Custom") — an empty start; you configure everything yourself. This is the default.

Recommendation: Choose the template that comes closest to your use case. It provides a sensibly worded system prompt as a starting point, which you then only need to adjust. "Custom" makes sense if your case doesn't fit any of the four categories.

After choosing a template, the configuration form appears with several tabs, grouped into "Configuration" and "Advanced". For the basics only two areas matter: the "Basics" tab (the "Basic settings" and "AI model" cards) and — for language — the "Persona" tab.

Note: Some tabs (e.g. "Knowledge", "FAQs", "Leads") are greyed out until the agent has been saved once. This is normal: these features need a saved agent ID. First create the agent with the basic details; afterwards all tabs are available.

Step 2: Basic settings ("Basic settings" card)

Here you define the agent's name and core behaviour.

  • Name (required field)
  • The agent's internal name, e.g. "Acme Customer Service". It helps you and your team find your way around the dashboard.

  • Default: empty. Recommendation: a short, unambiguous name per location/purpose. The field must not be empty, otherwise the agent can't be saved.

  • Description (optional)

  • A short note for your team ("Short description for your team"). It has no effect on the agent's behaviour.

  • Default: empty. Recommendation: fill it in as soon as you run several agents.

  • System prompt (required field — the heart of it all)

  • This is the single most important setting: the core instruction the agent follows on every response. Here you set out the role, task and behavioural rules in plain language (e.g. "You are a friendly customer service assistant for …"). If you've chosen a template, this field is already sensibly pre-filled.
  • Default: depending on the template, a suitable starter prompt; for "Custom", a neutral core instruction ("You are a helpful assistant. Be concise, friendly, and professional.").
  • Must be at least 10 characters long, otherwise an error appears.

  • Recommendation for good prompts:

    • Describe the role ("You are the assistant for the Restaurant Sonne").
    • State the tasks ("Answer questions about the menu, take orders step by step").
    • Set clear boundaries — especially for prices/facts, the formula "… use ONLY the knowledge base — never invent prices or products" is very effective in stopping the agent from making things up.
    • Keep it focused: an overloaded prompt waters down the responses. Detailed persona rules (tone of voice, forbidden topics) belong in the "Persona" tab, not in the prompt.

  • Welcome message ("Welcome", optional)

  • The first message that is automatically shown in the web chat widget when it opens (e.g. "Hello! How can I help?").
  • Default: "Hello! How can I help you today?" — adapt this to your language and brand.
  • Important (pitfall): SMS and WhatsApp do not use this welcome message. There, the first response is generated by the system prompt. An additional welcome message would duplicate the greeting there and cost an extra message segment — so the welcome message is deliberately intended for the web widget only.

Step 3: AI model and parameters ("AI model" card)

Here you choose the language model and set how the agent generates ("Choose a model and adjust the generation parameters").

  • Model ("Model")
  • Determines the underlying AI model. There are three options to choose from:
    • GPT-4o — "Recommended": the strongest all-round model, best response quality.
    • GPT-4o Mini — "Fast & affordable": noticeably cheaper and faster, somewhat less capable.
    • GPT-4 Turbo — "High performance": a powerful alternative.
  • Default: GPT-4o.
  • Recommendation: Stick with GPT-4o as long as you have no clear reason to switch. GPT-4o Mini is worth it for high message volumes and simple tasks (e.g. pure FAQ answering), to save costs.

  • Note: For internal helper tasks (e.g. short classifications in the background) Prezio automatically uses a fast, inexpensive model ("fast model") anyway — you don't need to configure anything for that. The selection above only concerns the model that writes the actual customer responses.

  • Max tokens ("Max. Tokens")

  • Limits the maximum length of a single response (tokens are roughly "word building blocks"). A higher value allows longer responses but can cost more; a lower value cuts off long responses.
  • Default: 2000. Permitted range in the form: 100 to 8000.

  • Recommendation: 2000 is suitable for most chat and support cases. Increase it only if responses regularly look cut off; lower it (e.g. to 800–1000) for deliberately concise responses, such as via SMS.

  • Temperature ("Temperature")

  • Controls how creative or how predictable the agent's wording is. The hint in the field reads: "0 = deterministic, 1 = creative". Low values give consistent, fact-faithful responses; high values give more variety and variance.
  • Default: 0.3. Permitted range: 0 to 1 (in steps of 0.1).

  • Recommendation: For support, booking, orders and anything fact-based, leave it at 0.2–0.4. Values above 0.7 only for a deliberately casual, "creative" marketing tone — unsuitable for factual information, because the risk of invented content rises.

  • Use knowledge base (RAG) — toggle ("kbToggle")

  • When active, the agent looks up relevant text excerpts ("chunks") from your knowledge base on every message and uses these for the response. This is the foundation for the agent actually knowing your menu, prices, opening hours and so on.
  • Default: on (enabled). Recommendation: Leave it switched on as soon as you upload content — otherwise the agent can only respond from the system prompt.

  • How it works: Retrieval works "recall-first". If nothing lies under the internal similarity threshold (cosine distance, default 0.5), the best near-match is used anyway (marked yellow as "recall-fallback" in the decision flow). Polluted, automatically crawled pages (navigation and boilerplate text) increase the distances and can cause good matches to be overlooked — in that case, re-upload clean text. You fine-tune the threshold later, conveniently, in the "AI Debug" section via the retrieval test with one-click save.

  • Top-K chunks ("Top-K Chunks")

  • Determines how many of the most relevant text excerpts feed into each response.
  • Default: 5. Permitted range: 1 to 20. The field is only operable when the knowledge base is switched on.

  • Recommendation: 5 is a good middle value. More chunks provide more context but cost more tokens and can dilute focus; fewer saves costs but risks missing information.

  • Daily budget (USD) ("Daily budget", visible to administrators only)

  • A cost cap per agent: as soon as the rolling 24-hour costs reach this mark, the public chat stops responding (it technically returns an "HTTP 429"). All provider costs are counted towards this.
  • Default: empty = no limit.
  • Recommendation: For publicly reachable agents, set a limit (e.g. 10.00) to prevent cost blow-ups from abuse or unusual demand. This field only appears for administrator accounts.

Step 4: Set the primary language ("Persona" tab)

You configure the agent's language not on the "Basics" tab, but under "Persona" in the "Languages" card.

  • Detect language automatically ("autoDetect") — toggle
  • When active, the agent responds in the language the customer writes in. When disabled, it always responds in the primary language.

  • Default: on. Recommendation: For a mixed audience, leave it switched on.

  • Primary language ("Primary language")

  • Choose from: Deutsch, English, Français, Español, Italiano, Nederlands, Português, Türkçe — or "(none — model decides)".
  • The meaning depends on the toggle above:
    • With automatic detection enabled, the primary language is the fallback if the customer's language isn't in your list of supported languages.
    • With detection disabled, the agent always responds in this language.
  • Default: "(none — model decides)".

  • Recommendation: Set the language of your main audience (e.g. German) as the primary language. Side note: English responses use a British accent and British spelling by default (for Voice).

  • Supported languages ("Supported languages", visible only when automatic detection is enabled)

  • Click the languages between which the agent may detect and switch. The primary language is always selected too (marked with ★) and can't be deselected.
  • Recommendation: Enable exactly the languages your customers actually speak — no more.

Saving

Click "Create agent" at the bottom. Prezio checks the required fields (name present, system prompt at least 10 characters). On errors, the form jumps back to the "Basics" tab and marks the affected fields red. After a successful save you land on the new agent's overview page — and all previously locked tabs (Knowledge, FAQs, Leads, etc.) are now unlocked.

Tips & pitfalls

  • The system prompt decides almost everything. Invest most of your time here. If the agent makes things up, add a clear rule such as "use only the knowledge base, do not invent prices/products".
  • Welcome message ≠ first response on every channel. The "welcome message" applies only to the web widget. SMS/WhatsApp greet via the system prompt — avoid a double greeting (it costs segments).
  • RAG without content achieves nothing. The "Use knowledge base" toggle is on by default, but the agent can only retrieve what you've actually uploaded. Maintain the knowledge base, otherwise the agent responds "flying blind".
  • Don't crank up the temperature. For factual information, leave it at 0.2–0.4. High temperature = more imagination = more risk of incorrect statements.
  • Max tokens limits the length, not the truth. If responses look cut off, increase the value (upper limit 8000); but that doesn't make responses "smarter".
  • Tabs only after saving. "Knowledge", "FAQs", "Leads" etc. only work after the first save — not an error, just the order of things.
  • Switching models is reversible at any time. Feel free to try GPT-4o Mini for cost tests; if the quality isn't enough, switch back to GPT-4o.
  • Daily budget for admins only. If you don't see the field, your account lacks the administrator role — contact the account owner.
  • Fine-tuning comes later. You analyse the knowledge-retrieval threshold and response quality not here, but in the "AI Debug" section (retrieval test with recommended threshold and one-click save, turn traces, eval and decision flow).