Rivet: Empower Your Chatbot With Semantic Routing! - no code tutorial
ฝัง
- เผยแพร่เมื่อ 2 ม.ค. 2024
- This video shows how to use semantic routing to work more efficiently with LLMs. It allows for:
- Blocking malicious/unwanted requests
- Automatic function calling
- and much more
Links:
- Get the example project: github.com/ai-made-approachab...
- Show some support to the github project: github.com/aurelio-labs/seman...
- Get Rivet: rivet.ironcladapp.com/
- Follow me on LinkedIn: / tim-k%c3%b6hler-ai-mad...
- If you enjoy my content, consider buying me a coffee: www.buymeacoffee.com/ai_made_... - วิทยาศาสตร์และเทคโนโลยี
I mentioned that I wanted to fix the small issue in the data import in the project before uploading it. But looks like it is a bug in Rivet that needs to be fixed there:
github.com/Ironclad/rivet/issues/294
Exact;y what I needed to know today.
this is awesome
Thx! And again big thanks to you for sharing the project and the idea!
Thanks for the interesting subject explanation. Really liked it.
Interesting.. I had exactly the same idea. Difference being that I actually used gpt-4 for semantic routing. I instructed it to only outpit the Route ID and nothing else. I then took the ID and route the conversation based on that. It works really well. It's amazing how many problems semantic routing can solve ..and it is currently not broadly used technique.
GPT-4 is probably simpler to do, but I would be worried about the cost. At least if you would run this as an application in production somewhere.
I saw a bunch of your videos now. Very interesting, but I see no explanation where the frontend for the users could be. Is there a video for this?
Currently trying to find a simple/low code solution. To sad Rivet is running on node.js. For python there is projects like gradio + streamlit, which have easy to use chat-components
@@AIMadeApproachable Well node.js is fine and could be easier to handle then streamlit. Fantastic would be a wordpress-plugin/compatible solution- so the Auth-stuff/subscriptions etc. could be handled separately by the most used CMS out there. This could be additionally a huge advantage over other systems. So far your solution is already the best in OSS.
very cool. I am wondering if this will work with different languages as my bots are multilingual. Maybe could be jailbroken by speaking another language. Will do some testing .....
I think you are absolutely right with your concerns. I am currently thinking how this should be setup for a production ready application and this is one of the topics I am thinking about. Probably will do a full video on this in the near future.
But for now: One solution can be to use a zero-shot-classification model, e.g. MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli
You can just give it "English" and "Other" as classifiers. And it is very reliable in detecting if the input is English or not. So you could run this first/or in parallel and only forward the user to the route/action once the language has been verified.