Massive intentional SPAM to the Chatbot

We have some malicious user intentionally spamming the chatbot constantly and they have spent I think over 200 dollars worth of tokens. This is rdicilous. We cant use the system like this(this is for a big client). WE NEED SOME WAY TO DETECT SPAM AND BAN THE USER FROM THE CHATBOT INSTANTLY!!!! HELP!!!

VoiceFlow team does not respond to emails at all.

VVossy We have some malicious user intentionally spamming the chatbot constantly and th...

Tico | Voiceflow Assistant•6/10/24, 4:22 PM

Voiceflow uses tokens as a form of currency to facilitate interactions with AI models. Depending on the subscription plan, users receive a specific monthly allocation of free tokens. For instance, the Starter Plan offers 100,000 tokens, the Pro Plan provides 2,000,000 tokens, and the Teams Plan includes 10,000,000 tokens. The Enterprise Plan offers a customized token quota to meet the needs of larger organizations.

Tokens play a crucial role in both input and output processes within Voiceflow. When interacting with AI models, tokens are expended based on the features utilized and the complexity of the interactions. For example, interactions involving question and answer sessions with the Knowledge Base feature may consume between 800 to 1500 tokens, while similar exchanges using the Response AI step with Memory may require anywhere from 150 to 2000 tokens.

Voiceflow is designed to support a high volume of users. It's built to scale and can easily handle thousands of users a day. However, it's always advisable to regularly monitor your application's performance to ensure it's operating at its best. With regards to the plan, keep in mind that each plan comes with a different amount of tokens which may be used up more quickly with a higher volume of users. So it is always advisable to upgrade your plan if you anticipate a huge increase in volume.

Please note that the information about handling spam and malicious users in Voiceflow chatbot is not supported by the source documents provided.

military-pinkOP•6/10/24, 4:51 PM

Anyone here??

military-pinkOP•6/10/24, 4:55 PM

Steve B. | IntelleSync.uk•6/10/24, 4:58 PM

Have a look at this solution from Cloudflare.

https://blog.cloudflare.com/firewall-for-ai

military-pinkOP•6/10/24, 5:04 PM

Thanks but it doesnt help. The spam protection layer needs to be implemented directly in VoiceFlow.

Steve B. | IntelleSync.uk•6/10/24, 5:28 PM

Yes but that isn't an option right now.

The Cloudflare solution would sit between Vf and the LLM via api, it's meant to detect malicious behaviour . I believe VF use this or something similar for Tico.

popular-magenta•6/10/24, 7:11 PM

Yea that is what we use for Tico - we dont have a spam filtering service built into VF at this time

popular-magenta•6/10/24, 7:12 PM

That is really rough though - send me your VF email, I'll add credit for more tokens to your account

W. Williams (SFT)•6/10/24, 8:35 PM

@Vossy Developing a good plan to protect your assistant is part of creating an assistant. You need to add checks to the assistant and handle bad actors. You can also implement restrictions:

Add checks to inputs to see if the input is valid.
Use the KB API to make sure the question is within the KB and handle bad responses.
Minimize LLM requests. I believe 60% of your bot should not use a LLM request.
Monitor (store) your requests to make sure they are valid and responses are correct.
Place restrictions in place based on the number of requests.
Use daily request limits (if you need to).
Possibly add filters to prevent specific content.
Use Clougflare AI Gateway like @Steve B. | IntelleSync.ai recommended.
Use LLM Caching.

Plus other things specific to your use case.

I like to call this BOT BULLETPROOFING. These are things that all devs do when creating any piece of software. I wish you the best of luck.