Loads your world in — and answers with sources.
Six file formats plus website crawl, one index. zep answers every question exclusively from the material you upload — and names the source it pulled the answer from. And when your sources say nothing about a question, that is exactly what it replies.
Six source types, one unified index.
No matter which tool your knowledge comes from — we unify, split and search it. You don't have to structure anything manually.
Multi-page documents and tables. Page numbers are preserved so every answer points to the right passage. Scanned PDFs need a text layer — pure image scans cannot be read.
Websites / URLs
The crawler starts at one address and follows internal links, parses main content and skips navigation, footers and cookie banners. robots.txt is respected.
Q&A and table lists
Question-answer pairs and price tables as CSV or JSON. Short entries stay their own section instead of getting lost in one large text block.
Markdown / TXT
Manuals, notes, docs straight from your repo. Markdown keeps its structure, plaintext is taken as-is — no conversion to PDF needed.
ZIP archives
One archive with mixed text files instead of twenty separate uploads. Exports from other tools (a Discord log as JSON, for example) can be ingested this way too.
Your own notes & feedback
Sharpen answers straight from the dashboard: correct a weak answer and turn that correction into a new knowledge source with one click.
Example · “How long does your delivery to Austria take?”
Standard shipping to Austria takes 3–4 business days from shipping confirmation. An express option has been available since January 2026 — the parcel then reaches you on the second business day, for a surcharge of €8.90. Both options are shown at checkout.
Three knowledge problems classic bots don't solve.
- Hallucinations. Classic LLM bots invent answers from their training material when they're unsure. In consulting, support or legal contexts that's damaging to business.
- A PDF is not an API. Manual, price table, website, exported notes — all structured differently. A chatbot has to unify these sources, split them and make them searchable.
- Knowledge goes stale. You change a pricing page, add an FAQ, add a shipping note — and the bot knows nothing about it and answers with last month's data.
Six tasks we handle for you.
Hybrid search, tenant isolation, GDPR, quality checks — sounds like an engineering sprint. At Zeptix it is already wired up; you never see these points as a task.
Hybrid search, not just vectors
Every question runs through embedding search, Postgres full-text search and a keyword path for numbers and prices at the same time. If you want, the bot setup lets you add cross-encoder re-ranking on top, which re-sorts the result list once more.
Strict tenant isolation
Your data lands in your workspace, never cross-indexed with other customers. Embeddings are tenant-scoped, no leakage between bots.
Indexing without downtime
While a new source is being ingested, the bot keeps answering from the previous state. As soon as indexing finishes it picks up the new material — no maintenance window.
Quality report per source
Every uploaded file gets a rating with concrete hints — for example when a PDF barely contains text because it is only scans. You notice at upload time that the bot cannot work with it, instead of at the first weak answer.
GDPR-compliant hosting
Embeddings and original files are stored in Germany (Hetzner). We provide a DPA under Art. 28 GDPR on request; when you delete a source, it disappears from the index along with its embeddings.
Marketplace as a knowledge pack
No material of your own? Activate a knowledge pack curated by Zeptix — it stays separate from your own sources and can be switched off again at any time.
FAQ on the knowledge base.
How exactly does RAG (Retrieval-Augmented Generation) work?
Before the LLM answers, we search your knowledge base for the passages most relevant to the question. These passages are sent to the LLM together with the question — and it answers exclusively based on that context. As a result, hallucinations statistically occur far less often.
Is my data used to train third-party models?
No. We use AI providers (Anthropic, Google, Groq, OpenAI) exclusively in inference mode with no training use. Our subprocessor list documents in full which provider is used for what. Your embeddings and original files stay in our database in Germany.
How large can my knowledge base be?
Knowledge storage depends on your plan and applies per bot: Starter 10 MB, Pro 50 MB, Business 200 MB. What counts is the size of the original files you upload. Once the quota is full, uploads are rejected — you then clean up old sources or switch plans. Plain text is very compact: 10 MB is several thousand pages.
Can I run a bot so that only my team can see it?
Yes — visibility is set per bot: public, team-only or private. Within a single bot there is currently no per-source grading: whatever you upload into a public bot can show up in its answers. Internal material therefore belongs in a separate, non-public bot.
How reliable are the source citations?
The bot names the passages it actually used for the answer — including the page number for PDFs. How many sources are pulled in and how strict the bot should be is configurable in the bot setup. If it finds no matching passage, it says so instead of guessing.
What happens when I delete a source?
It disappears from the index together with its embeddings and can no longer be used for answers. The deletion itself stays as an entry in the audit log so you can prove it later.
Pack in your knowledge. zep does the rest.
One PDF or one website crawl is enough — the bot is ready instantly and answers with source citations on the very first test.
