### Answer I don't have specific documentation about strategies for forcing the AI to consistently present exactly two time slots in all scenarios. However, based on community experiences, here are approaches that have helped with appointment booking reliability: Strategies to Improve Time Selection Consistency: 1. Use explicit step-by-step instructions in your prompt [(1) ]() . Structure your prompt with numbered steps that clearly define when and how to present time slots:

1. **Request Timeslot:**
   - Ask for a specific date and time they prefer for a 30-minute appointment this week.

2. **Check Availability:**
   - Check availability by calling function `check_availability_cal`.
   - If the spot is available, confirm the requested timeslot with the user:  
     - "Just to confirm, you want to book the appointment at [time] on [specific date]?"  
     - Ensure the chosen time is from the available slots.  
   - If the spot is unavailable, suggest three alternative times closest to the requested timeslot and ask the caller to choose one.  
   - Repeat the process until a valid timeslot is confirmed.

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

2. Pre-fill available appointment slots directly into the prompt before the call [(2) ]() . One community member found this approach worked better than real-time function calling, achieving roughly 90% accuracy by including available slots for approximately 1 week in advance . 3. Add a dedicated conversation node between availability checking and booking [(3) ]() . Structure your flow as: Check Availability Node → Conversation Node (where agent presents times) → User Confirmation → Book Appointment Node . In the conversation node, use a prompt like: "Tell the user all the available times found in the check_availability_cal function result. Wait for them to confirm which specific time they want before proceeding." 4. Include timezone and current time context [(4) ]() . Use

{{current_time}}

before triggering availability checks to help the LLM understand what "today" means . Known Limitation: Community members have reported that agents can be "very bad with dates" and function calling for appointment scheduling typically reaches "90-95% accuracy at best" [(2) ]() [(1) ]() . The inconsistency you're experiencing (9/10 tests working correctly) aligns with these reported accuracy levels.

​ To learn more, see the sources I considered: [[1] Discord message from @melvinglenford on 2025-03-06](https://discord.com/channels/1186424376300679209/1186462578310910112/1347206753783250995) [[2] Discord message from @.teamrick on 2024-05-06](https://discord.com/channels/1186424376300679209/1186462578310910112/1237050947147141162) [[3] Book Appointment Cal function returns error](https://community.retellai.com/t/book-appointment-cal-function-returns-error/278) [[4] Urgent: Cal.com Booked on different day - Potential Service Disruption with multiple issue.](https://discord.com/channels/1186424376300679209/1423174290458808381) ​

### Answer Based on the available sources, I can't find official documentation stating that 90% is the expected success rate for appointment booking with Retell AI. The 90% figure came from community experiences shared in Discord, not from Retell AI's official documentation [(1) ]() . What you can do to improve reliability: The official documentation suggests several configuration options that may help improve consistency: 1. Lower the temperature setting [(2) ]() Temperature controls randomness in responses. For appointment booking, use 0.1-0.3 for more deterministic behavior . Lower values make the LLM more consistent and focused . 2. Enable Structured Output Structured Output ensures the LLM strictly follows predefined schemas, which is particularly important for reliable function calling . When enabled, the model is constrained to output only valid function calls with all required parameters . This feature: - Eliminates missing or malformed function arguments - Prevents invalid function calls from being attempted - Ensures all outputs match expected schemas 3. Use Multi-Prompt structure instead of Single Prompt [(3) ]() For complex appointment booking flows, multi-prompt agents can help. As complexity increases, single prompt agents may experience behavioral drift and function calling issues . Multi-prompt agents provide: - Focused prompts for each conversation phase - State-specific functions only available when appropriate - Better function control Regarding your architect's comment: While context limits are a real consideration, the sources don't support the claim that you should accept a 10% error rate as inevitable. The documen

tation provides specific tools (temperature, structured output, multi-prompt) designed to improve reliability [(2) ]() [(3) ]() . I'd recommend trying structured output mode and lowering your temperature to 0.1-0.2 before accepting that 90% is the ceiling [(2) ]() .

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] Discord message from @ji350 on 2026-03-27](https://discord.com/channels/1186424376300679209/1186462578310910112/1486919318095794186) [[2] Configure LLM Options](https://docs.retellai.com/build/llm-options) [[3] Single/Multi Prompt Agent Overview](https://docs.retellai.com/build/single-multi-prompt/prompt-overview) ​

### Answer Based on your situation where temperature is minimized, structured output is enabled, and you've run 300+ tests, the official documentation doesn't provide additional configuration options beyond what you've already implemented [(1) ]() . What the documentation confirms: The official sources acknowledge that as complexity increases with appointment booking flows, agents may experience "behavioral drift" and "function calling issues" [(2) ]() . The documentation recommends keeping prompts concise, as "longer prompts can actually harm performance" [(3) ]() . Additional considerations from official guidance: 1. Fast Tier option [(1) ]() If not already enabled, Fast Tier provides 50% reduction in latency variance and 25% improvement in average response time . While primarily for latency, more consistent processing might help with function calling reliability . 2. Model selection [(3) ]() The documentation emphasizes checking "latency and throughput" when selecting models, noting that "time to first token + throughput of first sentence matters" . Different models may have varying function calling accuracy. 3. Prompt optimization The documentation states "when using function calling, set the temperature lower can help boost accuracy" , which you've done. It also suggests: "If you want to bound the agent behaviors, you can consider combining internal states (kind of like IVR tree) with different prompts & functions at different states" . The reality: The official documentation doesn't specify an expected error rate or guarantee 100% accuracy for function calling sce

narios. Given you've exhausted the documented optimization strategies (temperature, structured output, prompt refinement, multi-state structure) , the remaining ~10% inconsistency may represent the current practical limits of LLM-based function calling for complex appointment flows. If this error rate is unacceptable for your use case, you might need to explore custom validation layers or human-in-the-loop fallbacks, though these aren't detailed in the available documentation.

​ To learn more, see the sources I considered: [[1] Configure LLM Options](https://docs.retellai.com/build/llm-options) [[2] Single/Multi Prompt Agent Overview](https://docs.retellai.com/build/single-multi-prompt/prompt-overview) [[3] Custom LLM Best Practices](https://docs.retellai.com/integrate-llm/llm-best-practice) ​

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