We are developing voice-based conversations that generate an analysis plan containing structured data at the end of the call. When the conversation includes 4-8 questions, the structured data is generated correctly. However, when the number of questions increases to around 20, the analysis is still generated, but the structured JSON data is missing. Below are the call IDs: a8b4107f-4da7-4511-98b4-3464491359ac 200ba8c8-6b2c-4a9a-8e9c-5812cad7b0c3 4166bba0-1204-4942-b29a-82494b894d1c

Hi there! I've looked into your question about structured JSON data issues when handling a higher number of questions. While I couldn't find any explicit limitations on the number of questions in a voice-based conversation, there are several factors that might affect JSON data generation when scaling up: 1. **JSON Formatting Issues**: Ensure your JSON response is a single JSON object with a "message" key for the text and no unexpected nested structures. Incorrect formatting can prevent proper processing [Discord thread](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1327481613004701779). 2. **Incremental Data Handling**: With more questions, incremental transcriptions might lead to fragmented data collections [Discord thread](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1336360452208332831). 3. **Performance Impact**: Larger conversations might affect system performance, potentially impacting JSON generation [Discord thread](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1336243643043938395). For better results with many questions, consider these best practices: - Use consistent data structures with clear keys and values - Implement effective error handling and fallbacks - Adjust parameters like

numWordsToInterruptAssistant

for more stable conversations - Monitor performance regularly I couldn't find information specific to your call IDs or a known issue that occurs exactly at 20 questions. Would you be able to share more details about your implementation approach? Sources: - [Discord thread on JSON formatting](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1327481613004701779) - [Discord thread on incremental data handling](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1336360452208332831)

It is bound to happen cause as you increase the context window model will start hallucinating. You should do one thing after you got the conversation messages you should use higher LLM model to extract the details from it.

thanks