### Answer A 20-40% hallucination rate on dates is significant. Here are some strategies to mitigate this: Within Retell: 1. Use Structured Conversation Flows – Platforms that rely solely on open-ended prompting without control layers are more prone to drifting off course. [(1) ]() Conversation Flow provides a more constrained framework that helps minimize AI hallucinations with structured conversation pathways. [(2) ]() 2. Source-Grounded Prompting – Ensure agents only pull from verified, relevant knowledge bases. [(1) ]() 3. Add Validation Logic – Include checks for correct answers before confirming actions. 4. Lower Model Temperature – You can set

model_temperature

to a lower value (closer to 0) for more deterministic responses. Lower values are recommended for tool calling and structured outputs. [(3) ]() [(4) ]() 5. Clear Fallback Instructions – Include explicit fallback logic in your prompts like: "If the user didn't state this, leave it blank." [(5) ]() Consider External Handling: If date formatting is critical and hallucinations persist despite the above measures, handling date sentence formulation outside of Retell (via a custom tool or external API that returns pre-formatted date strings) can be a valid approach. This adds validation logic that checks for correct answers before confirming actions. [(1) ]() The key is ensuring the LLM doesn't have to "guess" or

generate dates—instead, it should retrieve verified data from your backend.

​ To learn more, see the sources I considered: [[1] 2025 B2B Guide to AI Phone Calls & Voice Agents | Retell AI | Retell AI](https://www.retellai.com/blog/b2b-guide-to-ai-phone-calls) [[2] Guide to AI Hallucinations & How to Fix Them | Retell AI](https://www.retellai.com/blog/the-ultimate-guide-to-ai-hallucinations-in-voice-agents-and-how-to-mitigate-them) [[3] List Retell LLMs](https://docs.retellai.com/api-references/list-retell-llms) [[4] Get Retell LLM](https://docs.retellai.com/api-references/get-retell-llm) [[5] Urgent: voice agent hallucinate with custom post call analysis!](https://discord.com/channels/1186424376300679209/1387483500441239673) ​

Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.

### Answer Since you're already using Conversation Flow with temperature at 0, here are additional strategies to address the date hallucination issue: Within Conversation Flow: 1. Split into smaller nodes – If a node contains instructions to collect multiple pieces of information (like dates) , the agent might inconsistently handle them. Splitting into separate nodes improves consistency. [(1) ]() 2. Change the node model – If instructions are concise but the agent struggles, try using a more capable LLM model for that specific node. 3. Add conversation finetune examples – To achieve a specific response style for dates, add finetune examples to help the model understand your expectations. External Handling Recommendation: Given you've already optimized the Conversation Flow settings and are still seeing 20-40% hallucination on dates, handling date sentence formulation externally is a reasonable approach. You can: - Use a custom function/tool that returns pre-formatted date strings from your backend - Have the LLM call this tool to retrieve verified date data rather than generating dates itself This ensures the LLM retrieves rather than generates date information, which eliminates the hallucination risk for that specific variable.

Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.

​ To learn more, see the sources I considered: [[1] Debug guide](https://docs.retellai.com/build/conversation-flow/debug-guide) ​