Hello, I have an issue where the LLM is stripping data before sending to the voice model (I think) based on the logs, I can see that the LLM has properly retrieved all the info needed, but then when sending the line to speak out loud to the voice model it ommited data which it generated with * BOLD* formatting Here's an extract of the log 11:38:49:673 [LOG] Model output: Parmi 11:38:49:673 [LOG] Model output: . 11:38:49:674 [LOG] Voice input: je vous recommande de considérer des marques reconnues pour leur qualité et durabilité. 11:38:49:699 [LOG] Model output: partenaires 11:38:49:699 [LOG] Model output: nos 11:38:49:718 [LOG] Model output: , 11:38:49:718 [LOG] Model output: ** 11:38:49:751 [LOG] Model output: Ak 11:38:49:751 [LOG] Model output: zo 11:38:49:777 [LOG] Model output: -N 11:38:49:777 [LOG] Model output: obel 11:38:49:800 [LOG] Model output: ** 11:38:49:806 [LOG] Model output: est 11:38:49:826 [LOG] Model output: excellente 11:38:49:826 [LOG] Model output: une 11:38:49:850 [LOG] Model output: option 11:38:49:851 [LOG] Model output: . 11:38:49:874 [LOG] Model output: Cette 11:38:49:875 [LOG] Voice input: Parmi nos partenaires, est une excellente option.


and it's missing the brand name "Akzo-Nobel" : 
 desired Voice input: Parmi nos partenaires Akzo-Nobel, est une excellente option.

hey i am having this issues as well

So - I managed to solve it... in a way. the issue is - when the Voice synth tries to pronounce what you see it stumbles on the * NAME * This is a bug for Vapi to solve, however you can solve it in the interip by forcing the LLM to not produce the BOLD / Italic and numbering. any form of prompting like - DO NOT use ** or forget Markdown doesn't work - the LLM produces them. so I tried another approach which was asking it to write out differently. I explained that the LLM output is sent directly to a voice synth - and is breaking the voice, so it should only produce pure text and formulate sentences which are complete and are to be spoke out. this forced the LLM to avoid using bold etc... altogether. I think the fact the LLM isn't aware that it's output isn't text is not helping but you can solve your problem by forcing it to remove the need for bold/bullets/etc.. and ask for clear text, but requesting a form of answering which demands clear text.

Hey Tancho, Whenever LLM generates the output, by default we apply some text formatting to it, which eventually removes the special characters' contents (such as any content inside the markdown symbols). These things are removed by default but the good part is you can toggle them as per your requirement. You can toggle different LLM utilities which are applied to the LLM output. I am dropping a [link](https://docs.google.com/document/d/1ZXeru4FxS1hEgltecipGl4Edvn4VC3whZdW8iob5dmU/edit?tab=t.0#heading=h.bk3gtu4bnekp)([LLM Voice Formatting Plan)](https://docs.google.com/document/d/1ZXeru4FxS1hEgltecipGl4Edvn4VC3whZdW8iob5dmU/edit?tab=t.0#heading=h.bk3gtu4bnekp) below where you can check out how to get things done.