How to Reset or Clean Contaminated Knowledge Bank Data for a Character

I’m working on a knowledge bank, and I’m encountering several issues. The responses generated by my character contain a lot of hallucinations. I implemented an anti-mixing prompt, which initially worked well. However, when I started testing another character, it seems like the original character’s knowledge bank got mixed up or contaminated.

Is there a way to reset or clean any previous information associated with my character to ensure it only uses its intended knowledge bank?

And most recently started to show us a pop up message saying “Server side error”

We have a professional licence and we are going to show it to our client next week.

Hello @design1,

Welcome to the Convai Developer Forum!

Could you please share the Character ID, the file name, example sessions, and the expected versus current responses? This will help us investigate the issue more effectively.

1 Like

Sorry, we initially thought the issue was related to the knowledge bank structure, but we’ve already tried all the suggestions we’ve found in the forum without success. I read that in some cases this might be related to a RAG check or an available update. Could you help us determine if the problem is connected to that?

Our Character ID is: 0e27c792-510a-11f0-8815-42010a7be01f

Example Sessions: 83d5d2cfe4ea3be341c8d02df803496e
This is mixing information from two different people.

Expected: OBREGÓN, Alejandro* (Barcelona, Spain, 1920 – Cartagena, Colombia, 1992)*
As a child, he lived between Spain and Colombia and traveled through the United Kingdom and the United States. After living in France for five years, he settled in Cartagena, where he worked with Guillermo Wiedemann, Santiago Martínez Delgado, Enrique Grau, and Ricardo Gómez Campuzano. Thanks to him, figurative expressionism gained great momentum in Colombia starting in the 1950s.

In 1958, he was awarded the first prize at the Hispano-American Biennial in Madrid and received numerous honors, including the Cruz de Rafael Núñez in Colombia. Among his most well-known works are Máscaras (1952), Homenaje a Zurbarán (1962), Flora (1966), Anunciata en verde (1970), Sortilegio de luna (1985), and La violencia, a piece that won the National Painting Prize at the 15th National Salon of Colombian Artists.

*He received the First Acquisition Prize for his work Torocóndor at the First American Art Biennial in 1962, which is part of the exhibition “Latin American Poetics in the IKA Collection (1958–1966).” or less

Current Respose: Alejandro Obregón was born in 1920 in Spain and moved to Colombia in 1924. He studied at the Art Students League in New York and at the Académie de la Grande Chaumière in Paris. He is considered one of the most important Colombian artists of the 20th century, known for his synthesis of European modernism and Colombian cultural elements. His work often explores themes of cultural identity and nature, using vibrant colors and dynamic compositions.

Hi @K3 , having the same issue, the character is not inventing data but it is mixing it, is the volume of data an issue? Accurate and precise answers based on the knowledge bank are needed :face_with_diagonal_mouth:

Hey @design1 @Matias_Juncos

I am Kamal from the LLM team, and I will help you in solving the issue. Could you just tell me a few things first.

  • What exactly do you mean by mixing? Are you saying, the character is getting context from a file which is not yet connected to it? I just want you to make sure that you connect only those files which your character needs.

If you are facing issues like the character’s knowledge getting contaminated from other files which are not connected to the character, then it must be a problem. Just confirm these once, and it would become easy for me to help you.

Thanks

Hey @kamal ! Thanks for your help!
Just to clarify — what I meant is that the character seems to be mixing up information from the same source file.

For example, with this session ID: a2a9343b9b2bdb63b0ac1c7822c28ac8 , the character should be retrieving data from a specific chunk, but instead, it pulls in some content from a different chunk related to another person’s biography and combining them. That’s the main issue we’re currently facing.

We’re preparing to present this to the client on Wednesday, so getting this fixed is quite important. Any ideas on how to prevent the model from blending content across different chunks?

Thanks again!

1 Like