1. Check that the project includes RAG
Open the project knowledge base. Your current plan must enable RAG and provide an available document quota. Starter does not include a knowledge base; check the project subscription page for the current limits and upgrade options.
Knowledge belongs to the project you add it to. Confirm the project and channel before importing, especially if you manage several products or customer inboxes.
2. Start with a small, clear FAQ
Choose questions customers repeatedly ask about setup, supported features and published policies. Put the answer next to each question and include exceptions. Remove conflicting or outdated versions before import.
The following is an example structure, not a statement of your product’s policy. Replace the bracketed fields with reviewed facts before using it.
Question: How do I change my plan? Answer: Open [actual settings page] and select [actual action]. When the change takes effect: [verified timing]. Exceptions: [limits and conditions]. If the change fails: [support contact or next step].
3. Add one source and check its processing
- File: the upload form accepts PDF, DOC, DOCX and TXT. Keep each file at or below the server’s 10 MiB upload limit.
- URL: provide a public webpage with the information you want to import. Start with one page. If using recursive crawling, review the depth, page limit and path scope before submitting.
- Text: paste a focused FAQ or short product reference directly into the text tab.
- After import: check the document’s processing status and available content before testing. A submitted upload is not proof that the knowledge is ready.
4. Test facts, wording and missing answers
- Ask a question with a clear answer in the source and compare the response with the original text.
- Ask the same question in different words to check whether the relevant information is still found.
- Ask a follow-up question in the same conversation and inspect whether the answer stays consistent.
- Ask something the source does not cover. Check that the AI does not invent prices, refund terms or product capabilities.
- Correct the source or project instructions when an answer fails, then run the same questions again.
Keep the source current
When a feature or policy changes, update the material in the project knowledge base and verify that the new content has been processed. Do not assume a previously imported URL always reflects later website edits.
Use your own approved support material. A knowledge base does not by itself connect the bot to live order data or grant it permission to perform account actions. Describe those boundaries in the project instructions.
Questions before you start
- Does the knowledge base train a new model?
- This feature uses retrieval-augmented generation (RAG): relevant source material can be retrieved for a reply. Adding a document is not the same as training or fine-tuning a new model.
- Why might a reply miss information I uploaded?
- Check the project, document processing state, source wording and plan limits. Also check whether the question is covered clearly in the source. Importing more documents alone does not guarantee a better reply.
- Can I use the same approach with Crisp and Chatwoot?
- Yes. Both use the GPT KeFu project knowledge base when the plan supports it. Their channel setup steps are different; follow the guide for your platform first.