Improve Knowledge Base Quality
A Zeptix bot is only as helpful as the knowledge you give it. The platform handles upload, processing, search, source display, and answer generation automatically. Your job is not to configure technology, but to prepare the content so the bot can understand and retrieve it unambiguously.
TL;DR
- Use text-based PDFs or clean website pages, not image screenshots.
- Split large topics into several small, clearly named files.
- Repeat product names, plan names, and key terms in every section.
- Write facts concretely: numbers, conditions, responsibilities, next steps.
- Test after every upload with real customer questions.
What Zeptix handles automatically
When you upload a PDF or index a website, Zeptix prepares the content for your chatbot. This includes, among other things:
- Extracting text from the sources.
- Splitting long content into meaningful sections.
- Finding relevant passages for a user question.
- Incorporating the found sources into the answer.
- Showing source references in the chat when they are available.
You do not need to build a technical pipeline or know any model parameters. The only thing that matters is: the clearer your content, the more reliably Zeptix finds the right passage.
Good sources are text-based
The most common quality mistake is PDFs that actually only contain images: scans, photos, exported flyers, or screenshots. People can read them, but a chatbot often gets barely any usable text out of them.
Check before uploading:
- Open the PDF.
- Try to select individual words with the mouse.
- If that does not work, OCR text recognition is probably needed.
- Then upload the text-based version.
For brochures, tables, or diagrams the rule is: add explanatory text. A diagram on its own rarely answers customer questions; a short section below it does.
Small files beat mega-PDFs
A file like unternehmen-komplett.pdf with prices, terms and conditions, team, FAQ, roadmap, and press is convenient for people, but often blurry for chatbots. Focused files are better:
preise-und-plaene.pdf
onboarding-erste-schritte.pdf
faq-kuendigung-und-rechnung.pdf
support-und-kontaktwege.pdf
produktfunktionen.pdf
This way Zeptix recognizes faster which source fits a question. It also makes the source display more understandable for your users.
Every section needs context
A good knowledge section works on its own too. Avoid sentences like:
It costs 49 euros and can be cancelled monthly.
Better:
The Acme Pro plan costs 49 euros per month and can be cancelled monthly. The Acme Pro plan includes 5 team members, invoice management, and priority support.
This sounds a bit more repetitive than marketing copy, but it helps the bot enormously: product name, plan name, price, cancellation, and features are all right next to each other.
Facts instead of marketing phrases
Marketing sentences like "a modern solution for demanding teams" rarely help. Your bot needs concrete information:
- Prices and contract terms.
- Limits and prerequisites.
- Links, email addresses, and responsibilities.
- Step-by-step procedures.
- Exceptions and special cases.
- What happens when something does not work.
Good rule of thumb: if a sentence cannot answer a concrete customer question, it belongs on the website rather than in the knowledge base.
Start website indexing deliberately
Website indexing is ideal if you already have good help pages, FAQ pages, or product pages. Do not start blindly with the entire website. Instead, first choose the pages that answer real support questions:
- Pricing and plan pages.
- Help and FAQ pages.
- Product documentation.
- Contact and support pages.
- Mandatory legal pages, if your bot should refer to them.
After indexing, you should test directly in the chat whether the most important questions are answered cleanly.
Test questions after every upload
Use five questions per source:
- A direct question whose answer appears verbatim in the source.
- A synonym question that customers would phrase differently.
- A comparison question, such as "What is the difference between Starter and Pro?"
- A boundary question that the bot should not answer.
- An action question, such as "What should I do now?"
If the bot answers vaguely, it usually is not "the AI being bad" but rather the source being too unclear, too large, or too promotional.