Chatbot with Own Data — Knowledge Base & RAG

A chatbot with your own data uses a tenant-specific knowledge base instead of relying only on general language capabilities. Zentor App supports three verified input methods: manual knowledge entries, website crawling and file upload. The knowledge base can contain FAQs, product descriptions, service instructions, internal process explanations and documents that customers frequently refer to. Supported file handling is limited to PDF, HTML, text and WordPress export content, with a maximum upload size of 100 MB.

The website crawler can process up to 300 pages in one run, is limited to eight minutes and respects `robots.txt`. A new crawl can be started after 24 hours. These limits matter when planning a large website import: a broad domain should be divided into relevant sections instead of assuming that every page will be processed in one pass. Pages that are blocked by `robots.txt`, unavailable or outside the configured scope cannot become reliable sources for the chatbot.

Retrieval-Augmented Generation, or RAG, searches the tenant’s knowledge base for relevant material and supplies that context to the language model. RAG is disabled by default and must be activated for the tenant. Uploading documents alone does not guarantee that they are already used in answers. Administrators should check the processing status and test the bot after activation. The verified package facts provide 1,000 MB of knowledge-base storage and 100 knowledge entries for both Starter and Individual (Individuell).

Good source quality is more important than volume. Conflicting prices, outdated service rules and duplicated documents can lead to weak retrieval or contradictory responses. Each source should have a clear purpose and a current owner. After a product or policy changes, the affected entry should be updated or removed rather than leaving both versions active.

Testing should include direct questions, paraphrased questions and cases where the answer is absent. The bot should either state that the information is unavailable or use the configured handover path instead of inventing a confident answer. Tenant isolation remains enforced on the server side, so one tenant’s documents must not be used for another tenant’s conversation.