Paulo C.

I mean, this is super frustrating. This is part of the prompt I'm using: "If the call is answered by a voicemail, phone directory, automated system, or any kind of interactive response system (such as "press 1 for sales"), you must immediately hang up the phone. Do not speak. Do not leave a message. Do not acknowledge the system. Do not say anything about ending the call. Just hang up immediately and silently. Never speak your internal thoughts, actions, or reasoning out loud — for example, do not say things like "I'm recognizing this is a voicemail, so I will hang up." You must behave naturally, like a human caller, and only speak directly to real people." I can't be more clear than that. We tried different prompts using different words to say all of the things that the bot is supposed to do. The image is showing what the bot actually did. 27 seconds of the bot doing everything that it was supposed to not be doing

Actually, I realized that I'm having this issues with a similar flow that's already using Claude 4 Sonnet for the calls. When the call starts, the first delay after the person says "hi" can be from 3 or 4 seconds sometimes. Since the person don't realize that they're talking to an AI at first, they say something like "hey is anyone there?" because they don't get an instant answer because of the delay and then the AI have to recalculate what it was going to say for the first "hi" that the person said. This may cause the AI to "glitch", because the text it was supposed to say appears on the transcriptions in the tasks but the AI don't actually say all of that to the person on the call

Thanks, Alon! I'll give it a try!

Hi guys, this is an example https://chat.lindy.ai/griffin-morriss-workspace/lindy/mkt-call-leads-mid-score-v2-68223cb4083cb1540e09f26c/tasks?task=68628d0a6bd57e73cdc2e9d3

Hi Alon! Sorry for taking so long to respond. I missed your response. Here's the link: https://chat.lindy.ai/griffin-morriss-workspace/lindy/mkt-call-leads-low-score-v2-682cc31545551db50380abc2/tasks?task=685ee1169646483e629d0885

Forgot to ask something but if I build it the way you recommended, the model I should use for the perplexity step that is only going to do the research is Sonar or Sonar pro, correct? Because the score is being given by the ai agent, so the perplexity step isn't doing any reasoning

The Perplexity step uses a text input that’s set to AI prompt mode. In this prompt, we give instructions based on the information received in the previous step, define what it should look for, and then provide the “rules” it should follow to assign a score based on what it finds. Given that it's in AI prompt mode and using GPT-4o Mini, does that change your answer in any way? Either way, I’ll test the approach you recommended by letting Perplexity handle the research and assigning the score in the next step. Thanks so much!

Yes, that’s helpful! Thank you! Just to confirm: we’re using the query field to “teach” this Perplexity step the rules it should follow to assign scores based on the information it finds. Based on the screenshots I uploaded, the scoring is being handled by the GPT-4o Mini model, and the actual research is being done by the Sonar Reasoning model. Is that correct? And secondly, are these the best models for this use case? With the reasoning for the score being done by GPT-4o Mini and the research being handled by Sonar Pro instead of Sonar Reasoning, for example?