How Does a Zeptix Bot Work?
Anyone who has used ChatGPT often thinks: "My bot is a specialized AI that was trained on my data." That is wrong — and this exact assumption leads to wrong expectations. Once you understand how a Zeptix bot really works, you immediately write better knowledge bases and better system prompts.
TL;DR — The three stages of a Zeptix bot
A Zeptix bot is not a trained AI. It is a three-stage system:
| Stage | What happens | What you control |
|---|---|---|
| 1. Personality (system prompt) | An instruction to the AI model about who the bot is, how it sounds, what it declines | Dashboard tab "System prompt" |
| 2. Knowledge | For every question, Zeptix searches for fitting passages from your sources | Dashboard tab "Knowledge base" |
| 3. Answer logic | Zeptix generates an answer from this, considers protection rules and shows sources when available | Dashboard tab "Model" and plan functions |
Important: Your data is NOT used for training. It is only searched at runtime (retrieval), not built into a model update. This has two consequences:
- Privacy: Your content never flows into a future model update.
- Leverage: You cannot accumulate "training data improvements" over time. Every answer is created fresh from the current state of your content + system prompt.
→ From this it follows: Quality comes from a clean system prompt + cleanly structured PDFs, not from volume.
Stage 1 — Personality (system prompt)
The system prompt is a text instruction to the model that is sent along before every answer. It defines:
- Identity: Who is the bot? ("You are Training-Tom, the AI coach for hobby strength athletes.")
- Domain: What is it for? ("You answer training questions about the FitnessHub programs.")
- Tone: How does it sound? ("Casual, informal, motivating.")
- Boundaries: What does it decline? ("No diet plans, no training tips for those under 18.")
A good system prompt is the most important quality lever — a well-written personality can deliver brilliant answers with a minimal knowledge base; a poor personality makes even a perfect knowledge base boring.
Detailed guide → Writing a system prompt.
Stage 2 — Knowledge
Instead of training a model on your data, Zeptix searches for fitting passages in your uploaded content for every question. These sources help the bot answer concretely and traceably.
In simple terms
- You upload a PDF or index website content.
- Zeptix prepares the text so that it can be found later.
- A user asks a question.
- Zeptix searches for fitting passages from your knowledge base.
- The bot formulates an answer from this in your desired tone.
- When sources are available, they are shown in the chat.
What does this mean for you?
This means for your content:
- Large collection files are worse than several focused files.
- Short sections with clear headings are found more reliably.
- Tables need context, otherwise the bot understands individual rows less well.
- Price comparisons, cancellation rules and support paths should each be explained compactly and completely.
Why key terms must be repeated
Zeptix finds content more reliably when important terms are repeated in the relevant section. Anyone who names a concept once and afterward only refers to it with "it", "the system" or "the solution" makes the source unnecessarily fuzzy.
Bad (vague reference):
The system is controlled via the web interface. It offers all the necessary functions. The operation is intuitive.
Good (key term repeated):
Acme Pro is controlled via the web interface app.acme.com. Acme Pro offers dashboard, reporting, team management and billing in one interface. The operation of Acme Pro is optimized for mouse + keyboard.
More on this → Splitting the knowledge base correctly.
Stage 3 — Automatic answer logic
Zeptix chooses the fitting answer logic for your plan and the respective request in the background. You do not have to pick technical models for every question.
What you can control as an owner
- Default: A plan-aware router automatically chooses per request.
- Override (in the dashboard tab "Model"): you can force your bot fixedly onto the large reasoning model of your plan tier — sensible if your bot almost only gets complex questions.
- Own API keys (Business only): you store your own provider keys. Your traffic then goes directly to your provider account, Zeptix only manages the routing.
What you cannot control
- Internal model or infrastructure details. Your bot should help users, not explain technical platform details.
- Low-level technical model parameters. Zeptix sets production-safe defaults so that bots stay stable.
How it all works together — a concrete visitor request
Let's say a visitor asks your acme-buddy.zeptix.io bot the question "How do I cancel my Pro subscription?":
- Zeptix checks access, limits and the bot's context.
- Zeptix searches for fitting passages from your knowledge base.
- Personality, tone and sources are combined.
- The answer is streamed live into the chat.
- Sources are shown when fitting sources were used.
- The usage is booked as credits.
All of this happens in 2–8 seconds, depending on model load and answer length.
What you have understood now
If you have read this article, you can answer the following questions:
- Why does it make a difference whether you split PDFs or deliver them in a single mega file? → Focused files deliver clearer sources.
- Why do key-term repetitions boost the hit rate? → The bot can assign fitting sections more unambiguously.
- Why is it pointless to write model instructions into PDFs? → The knowledge base is for facts, not for behavior.
- Why should your bot not claim to be a human? → Transparency protects you and your users.
Where to read next
- Improving knowledge base quality
- Steering answer behavior and boundaries
- Prepare your first knowledge PDF correctly — the practice for your first file.