RAG Intents
Hi - I just watched this video from Voiceflow YouTube channel about LLM+RAG Intents: https://www.youtube.com/watch?v=bavOGpeCC3M
But in nowadays dashboard I don't see that setting anymore on the "Intents" page where I can pick LLM+RAG?
Can someone please help? Thank you
17 Replies
Hey there!
The option labelled “LLM + RAG” in the video has since been rolled into our new LLM-powered (Hybrid) Intent Classification engine, so you won’t see a separate toggle any more. How it works now 🔧 • The classifier first checks your regular NLU model.
• It then sends the top candidate intents (plus their descriptions) to an LLM.
• That LLM uses retrieval over those descriptions (the “RAG” part) to pick the best match.
Because the RAG step is built-in, there’s no extra switch to enable. Where to turn it on ✨ 1. Open your project → Intents tab.
2. Click the gear icon (or “⋮” > Settings) in the top-right of the Intent CMS.
3. Enable LLM Intent Classification and pick a model / temperature if you wish. That’s it—once enabled, every Detect Intent step will automatically use the hybrid NLU + LLM (RAG) flow shown in the video. Reference doc: https://docs.voiceflow.com/docs/intent-classification-using-llms-hybrid
Hope that clears things up! 🙌
The option labelled “LLM + RAG” in the video has since been rolled into our new LLM-powered (Hybrid) Intent Classification engine, so you won’t see a separate toggle any more. How it works now 🔧 • The classifier first checks your regular NLU model.
• It then sends the top candidate intents (plus their descriptions) to an LLM.
• That LLM uses retrieval over those descriptions (the “RAG” part) to pick the best match.
Because the RAG step is built-in, there’s no extra switch to enable. Where to turn it on ✨ 1. Open your project → Intents tab.
2. Click the gear icon (or “⋮” > Settings) in the top-right of the Intent CMS.
3. Enable LLM Intent Classification and pick a model / temperature if you wish. That’s it—once enabled, every Detect Intent step will automatically use the hybrid NLU + LLM (RAG) flow shown in the video. Reference doc: https://docs.voiceflow.com/docs/intent-classification-using-llms-hybrid
Hope that clears things up! 🙌
@Tico | Voiceflow Assistant There is no gear icon in top right of the "Intents" tab, here is my screenshot:
The only setting option I see is this:
But there's no button to "Enable LLM Intent Classification".
Also, what would I put in the prompt here?
1 more thing, the reference doc you linked isn't a valid page anymore... https://docs.voiceflow.com/docs/intent-classification-using-llms-hybrid
@Michael Hood (Head of Product) Do you know maybe? Thank you
Still trying to figure this one out… @Moderator can you help please?
It is in your agent under content > Intents > Settings
Thank you @W. Williams (SFT) 🙏
Would you recommend always using the RAG + LLM setting?
Also do you have any resources on how to craft that “prompt” for intent classification setting?
hey @Marlon
The other options are deprecated, the only option is RAG + LLM, hence you can’t find the toggles any more.
Ohhh thanks @San , so then I don't need to make the prompt for intent classification setting? Or still need to?
Still need to, but the you should get a default prompt.
Oh gotcha. Do you maybe have any resources on crafting these prompts for intent classification? Anything that I can give to an LLM so it understands how to write it for me @San thanks a ton man!
I think you refereeing to intents and utterances rather than this prompt.
Make sure you create a good description for intent as LLM will use that to make the final decision.
Try to create well separated intents without overlapping utterances.
We have also generate utterances for an intent, you can try to utilize that + polish after if needed.
Gotcha. Thank you @San I appreciate you helping me man
Sure, happy to help.
i have the same problem, i was using the naturla language understanding but right now i also had to switch to RAG + LLM. my chatbot was working fine before, understanding the intents + utterances, since ive worked very precisely on them. the questions that didnt match any of these had a fallback, that would still answer according to the prompt that was triggered. now, my chatbot cannot handle the questions that are not part of an intent/utterance. so ... it just assumes that the question fits an intent and gives an answer that isnt good ... what do i do. you said to "create a good description for intents as LLM" what do you mean?
Switching to RAG can require tuning the utterances again as the underlying technics are different.
About the description, intents have name and description, that description is used for LLMs to understand which one to pick (similar to tools for agents).
ahhh thank you so much!!! this is super helpful