Thank you Ibrahim that was a comprehensive and very useful video. I love the 4mn padding!! America where time is money!! In our part of the world it would be: after lunch, or after Xmas 2025!! If you're lucky... and don't stop off for a coffee at their friend's house before arriving at the destination. It's interesting that RetellAI looks like Voiceflow or other chatbots. It's only question of time before the technology of chatbot and voicebot merge completely and the companies running them merge too. I look forward to your next video. The only short coming of Retell seems the limitation in languages. ElevenLabs provides 22 and Retell only a dozen. Why?
Assalam alaykum brother, really appreaciate your effort on this video! Just a question. Would it be possible to have a node behave as both: 1) a global node, and.. 2) a transitioned node ? I guess im trying to figure out a solution if in case the conversation flow I built is not 100% reflecting real use case sceanios (since we only link transitions to expected conversation nodes), while real life users can be much more random/ unexpected. Setting a node to 'global' would enable it to be 'reached' (if necessary) regardless of the defined transition flows. Wondering if this configuration is something you have also considered.
Hey Ibrahim, You stated multistate agents are inferior to conversation flows for high levels of control, as does the Retell documentation. However, after using both, I'm not sure what mechanisms of conversational flows give more control. It doesn't seem to offer better state management. The edges have less control,. as you can't specify an output json model. It almost seems like multi-prompt repackaged with a new UI. Am I missed something? Looking forward to your insight Thanks!
Great question. Ultimately, the main advantage with flows is granular control over the agent. I'll give some examples: With flows, you can have the agent say a sentence the exact same way you'd like (static sentences in conversation nodes), something that's not possible except with the first message in multi-prompt agents. Conversation nodes should also not be thought of states in multi-prompt -- conversation nodes are typically smaller versions of these nodes. They're not a whole prompt, but rather much shorter. You'll probably notice that as a result, flows are made of much more nodes than multi-prompt agents are with states. Multi-prompt agents couldn't really do this, otherwise there'd be less performance because they don't perform as well when there's many states/transitions. One thing also is that you can specify exactly when a call transfer/function node should execute. Before, these would have to be embedded in states. But now they can be like just a normal node, with their own transitions and such. There's much more advantages (like some of the settings such as "no response", etc.), but I'd suggest going through example flows and spending more time playing with them, and I think the difference will become more evident. Hope this clears up some of the confusion!
@@theibrahimmunir Thank you for the detailed response! So it seems like conversational flows can more effectively limit the state? That's great. I've been working more with multi-prompt and it does seem like the state management is pretty loose. Variables extracted in previous prompts are still available, and previous prompts still seem to have full influence. Going to try to implement my current conversation in Flows and pressure test :) I'm really curious about how multi-prompt and Flows are implemented on the Retell backend. I think this would give users a better idea of how to use them.. but there isn't much documentation available it seems. If you have any insight on this, I think it could be an excellent future video topic!
In my opinion, an ideal node should only have context from the overall model context + variables explicitly passed from an input node. I feel like it's not clear how state is passed between nodes in a Flow. Hopefully they can add this in the future, as it's only in beta.
I don't have a way to share the flow (not that I know of), but you can still copy the arrangement and prompts or do something similar. Hope this helps.
Hey Joe, yes call transfers do work. In my example call, since I was talking to the agent on the dashboard and not through a real phone number, it just hanged up. But in the case of real phone call, it would transfer the call to number you specify the call transfer node.
Assalamu Alikom brother I want to open my AI agency and need some help or guidance there is a lot of information that actually gets you frustrated is it possible for me to book a call with you and give me some advice or consultation to put me on the right way
Does this not restrict the Ai? Is it not better for it to be a more free? Ai is getting better and better all the time, so to make it go down a path rather than be free I think we are restricting its full potential. Voice Ai is much better than chatbots so why are we trying to put voice ai into a chatbot box?
It comes down to the use case you have. There's pros and cons for each approach, but the people who will appreciate these flows the most are enterprises/B2B use cases, and that's usually the main target for these voice AI providers. Enterprise/B2B calls are typically more transactional. There's a couple of possible outcomes for the call, and the goal of the call is to keep it as short as possible, professional, and achieve one of those outcomes. And a lot of times with enterprise/B2B, there's a huge penalty (e.g. monetary penalty for example) if the goal or outcome desired is not reached). So these companies want as much control with the agents to make sure the goal or outcome is reached. One last thing is that the conventional single-prompt agents are just simply a system prompt. This means you have to dump everything in that prompt, with exception to the knowledge base. There's a lot of research out there showing how LLMs struggle more as you continue to add more tokens/context to the prompt. So agent builders such as flows attack this problem by still using this system prompt, but shrinking the number of tokens/context under the hood. It only introduces information about how to qualify a lead into the system prompt, when the agent hits that node. And then as soon as it leaves that node, that info is removed from the prompt. That way, the agent is always focused only at what matter at that specific point of the call -- there's less tokens in the prompt and better performance as a result. Pretty similar to us humans -- if you throw a bunch of information at us, we'll struggle to remember even pieces of it. But if you tell us step by step, we're more likely to remember it all. Hope this helps.
Can you please make video of batch outbound calling agent using the same technologies? For example using google scraper scan google maps to get list of my potential clients and call each а them using AI to offer my service + send them emails. Then send me results via sms on my cellphone(don't see it quite often using sms as tool) about each conversation result. Hope you like this idea, would much appreciate it🙏
That was an amazing video. Hope to see more like this very soon.
Thanks Ibrahim!
Bro the voice is so clean! When will French voices be that smooth 😮💨
Thanks man!
By the way you're very good teacher 👍
Thanks Vladimir!
Great tutorial on Voice Agents. thanks for sharing
Thanks Gurinder!
Thank you @Ibrahim!
Thanks!
perfect video...
My man Zehaan, appreciate it.
Thank you Ibrahim that was a comprehensive and very useful video. I love the 4mn padding!! America where time is money!! In our part of the world it would be: after lunch, or after Xmas 2025!! If you're lucky... and don't stop off for a coffee at their friend's house before arriving at the destination. It's interesting that RetellAI looks like Voiceflow or other chatbots. It's only question of time before the technology of chatbot and voicebot merge completely and the companies running them merge too. I look forward to your next video. The only short coming of Retell seems the limitation in languages. ElevenLabs provides 22 and Retell only a dozen. Why?
Assalam alaykum brother, really appreaciate your effort on this video!
Just a question. Would it be possible to have a node behave as both:
1) a global node, and..
2) a transitioned node ?
I guess im trying to figure out a solution if in case the conversation flow I built is not 100% reflecting real use case sceanios (since we only link transitions to expected conversation nodes), while real life users can be much more random/ unexpected.
Setting a node to 'global' would enable it to be 'reached' (if necessary) regardless of the defined transition flows.
Wondering if this configuration is something you have also considered.
Hey Ibrahim,
You stated multistate agents are inferior to conversation flows for high levels of control, as does the Retell documentation. However, after using both, I'm not sure what mechanisms of conversational flows give more control. It doesn't seem to offer better state management. The edges have less control,. as you can't specify an output json model. It almost seems like multi-prompt repackaged with a new UI.
Am I missed something? Looking forward to your insight
Thanks!
I'm not sure if it's a advantage or a disadvantage, but tool prompts in Flows don't allow you to add context for calling the tool.
Great question. Ultimately, the main advantage with flows is granular control over the agent. I'll give some examples:
With flows, you can have the agent say a sentence the exact same way you'd like (static sentences in conversation nodes), something that's not possible except with the first message in multi-prompt agents. Conversation nodes should also not be thought of states in multi-prompt -- conversation nodes are typically smaller versions of these nodes. They're not a whole prompt, but rather much shorter. You'll probably notice that as a result, flows are made of much more nodes than multi-prompt agents are with states. Multi-prompt agents couldn't really do this, otherwise there'd be less performance because they don't perform as well when there's many states/transitions.
One thing also is that you can specify exactly when a call transfer/function node should execute. Before, these would have to be embedded in states. But now they can be like just a normal node, with their own transitions and such. There's much more advantages (like some of the settings such as "no response", etc.), but I'd suggest going through example flows and spending more time playing with them, and I think the difference will become more evident.
Hope this clears up some of the confusion!
@@theibrahimmunir Thank you for the detailed response!
So it seems like conversational flows can more effectively limit the state? That's great.
I've been working more with multi-prompt and it does seem like the state management is pretty loose. Variables extracted in previous prompts are still available, and previous prompts still seem to have full influence.
Going to try to implement my current conversation in Flows and pressure test :)
I'm really curious about how multi-prompt and Flows are implemented on the Retell backend. I think this would give users a better idea of how to use them.. but there isn't much documentation available it seems. If you have any insight on this, I think it could be an excellent future video topic!
In my opinion, an ideal node should only have context from the overall model context + variables explicitly passed from an input node. I feel like it's not clear how state is passed between nodes in a Flow. Hopefully they can add this in the future, as it's only in beta.
Thank you for video, do you have template of your voice agent shown in video🙏
I don't have a way to share the flow (not that I know of), but you can still copy the arrangement and prompts or do something similar. Hope this helps.
So how would you sell a retell ai agent? Do you simply update the account to their business info after sale?
Does phone transfer work as well?
Hey Joe, yes call transfers do work. In my example call, since I was talking to the agent on the dashboard and not through a real phone number, it just hanged up. But in the case of real phone call, it would transfer the call to number you specify the call transfer node.
Was that a cloned (custom) voice?
Hey Janysh! That was a custom 11Labs voice.
Assalamu Alikom brother I want to open my AI agency and need some help or guidance there is a lot of information that actually gets you frustrated is it possible for me to book a call with you and give me some advice or consultation to put me on the right way
Walaikum Salam, feel free to email me at ibrahim@shadow-ai.co and I can try to help
Make one for real estate lead qualification
Does this not restrict the Ai? Is it not better for it to be a more free? Ai is getting better and better all the time, so to make it go down a path rather than be free I think we are restricting its full potential. Voice Ai is much better than chatbots so why are we trying to put voice ai into a chatbot box?
It comes down to the use case you have. There's pros and cons for each approach, but the people who will appreciate these flows the most are enterprises/B2B use cases, and that's usually the main target for these voice AI providers.
Enterprise/B2B calls are typically more transactional. There's a couple of possible outcomes for the call, and the goal of the call is to keep it as short as possible, professional, and achieve one of those outcomes. And a lot of times with enterprise/B2B, there's a huge penalty (e.g. monetary penalty for example) if the goal or outcome desired is not reached). So these companies want as much control with the agents to make sure the goal or outcome is reached.
One last thing is that the conventional single-prompt agents are just simply a system prompt. This means you have to dump everything in that prompt, with exception to the knowledge base. There's a lot of research out there showing how LLMs struggle more as you continue to add more tokens/context to the prompt. So agent builders such as flows attack this problem by still using this system prompt, but shrinking the number of tokens/context under the hood. It only introduces information about how to qualify a lead into the system prompt, when the agent hits that node. And then as soon as it leaves that node, that info is removed from the prompt. That way, the agent is always focused only at what matter at that specific point of the call -- there's less tokens in the prompt and better performance as a result.
Pretty similar to us humans -- if you throw a bunch of information at us, we'll struggle to remember even pieces of it. But if you tell us step by step, we're more likely to remember it all.
Hope this helps.
@@theibrahimmunir great explanation brother. Thank you.
Can you please make video of batch outbound calling agent using the same technologies? For example using google scraper scan google maps to get list of my potential clients and call each а them using AI to offer my service + send them emails. Then send me results via sms on my cellphone(don't see it quite often using sms as tool) about each conversation result. Hope you like this idea, would much appreciate it🙏