Chatbots (Flowise) & RAG (Supabase V. DB) - n8n upsert [SOLVED]

Hey everyone

I've developed a chatbot using Flowise, and integrated it with my Supabase database.

Through n8n, I'm upserting a single .docx document into the Supabase vector store. However, when querying the database via the chatbot, I'm encountering hallucinations—responses that are inaccurate and potentially misleading, which is unacceptable for the chatbot's reliability.

I've configured my chatbot with a temperature setting of 0.1 and included directives in the system prompt such as "Always use KB for answers" and "Never write anything you don't know for sure."

Despite these precautions, I'm still encountering hallucinations—responses that are inaccurate and potentially misleading.

I'm seeking advice on how to adjust my chatbot to prevent it from generating such content, without having to enumerate every possible "do not answer..." scenario in the system prompt.

It's weird — whenever I ask the chatbot about pricing for regular customers versus businesses, it sometimes responds with, "I don't know anything about prices—please contact us." This is unexpected, especially since there's a dedicated section on pricing in the knowledge base.

Are there strategies or best practices to achieve this?

23 comments