### Answer Retell AI blocks calls to and from the following countries regardless of your configuration: [(1) ]() - Cuba (CU) - Iran (IR) - North Korea (KP) - Syria (SY) - Russia (RU) - Belarus (BY) - Venezuela (VE) Calls to or from these sanctioned countries will be automatically rejected. Additionally, for SMS specifically, the service currently supports USA numbers only. [(2) ]()

​ To learn more, see the sources I considered: [[1] Fraud Protection](https://docs.retellai.com/reliability/fraud-protection) [[2] TEXT MESSAGE ISSUE](https://discord.com/channels/1186424376300679209/1424595052595904542) ​

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

Thanks for the follow-up, @Blekk. I'll get back to you soon.

:** In your agent's security and fallback settings, under Personal Info Redaction, you can select which PII types to automatically redact. [(1) ]() [(2) ]() Administrators and developers can still view redacted information when needed for auditing purposes.

​ To learn more, see the sources I considered: [[1] Secure AI Conversations with Retell AI PII Redaction Tool | Retell AI | Retell AI](https://www.retellai.com/blog/introducing-retell-ai-pii-redaction-data-security-made-easy) [[2] Retell PII Redaction | Remove personal info from audio and transcription]() [[3] Compliance](https://www.retellai.com/glossary/compliance) [[4] 5 Enterprise AI Call Rollout Mistakes to Avoid | Retell AI | Retell AI](https://www.retellai.com/blog/the-5-most-costly-mistakes-enterprises-make-with-ai-call-rollouts-how-to-recover) [[5] Conversational AI in Insurance: 2026 Guide | Retell AI](https://www.retellai.com/blog/conversational-ai-in-insurance) [[6] Hipaa Compliance](https://discord.com/channels/1186424376300679209/1410647478537293905) ​

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

### Answer Based on the available sources, I can provide some information about data storage and internet access: Data Storage: When you pass data through dynamic variables or metadata, these values are stored in plaintext with every call record. [(1) ]() Both

dv

(dynamic variables) and

metadata

values are visible in call logs and API responses. The sources explicitly warn: "Do not use them to pass API keys, database credentials, or other sensitive secrets." Internet Access: Retell AI's Code Tool feature does provide internet access through the

fetch()

function, which allows making HTTP requests to external APIs. The documentation notes: "Enabling outbound HTTP requests from an LLM-invoked tool increases security and operational risk." For security-sensitive integrations that require accessing internal systems, databases, or handling sensitive credentials, the sources recommend using Custom Functions hosted on your own backend instead of Code Tool. Important Security Note: The sources emphasize that dynamic variables and metadata "are not designed for secret management." For workflows requiring sensitive credentials, authentication, audit logging, or write access to production systems, you should use Custom Functions on your own infrastructure rather than passing sensitive data through Retell's platform. I don't have specific information in the available sources about other storage locations beyond call logs for CRM data. For detailed information about data storage architecture and security practices, you may want to check Retell AI's security documentation or contact their support team directly.

​ To learn more, see the sources I considered: [[1] Code Tool](https://docs.retellai.com/build/single-multi-prompt/code-tool) ​

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

### Answer I don't have specific information in the available sources about "Average tokens from prior user/assistant turns and tool call/result messages carried forward in each LLM request" or detailed metrics about token usage patterns. However, the sources do mention token-based pricing changes. As of June 1st, Retell AI implemented token-based LLM pricing where prompts up to 3,500 tokens follow the existing flat rate, but prompts exceeding 3,500 tokens scale proportionally in cost. [(1) ]() For example, with OpenAI GPT-4.1: - A prompt with 3,200 tokens is billed at the base rate - A prompt with 4,500 tokens is billed at 4500 / 3500 × base rate The sources also note that Retell LLM has a token length limit of maximum 8192 tokens. For multi-state prompts, this includes the longest state plus the general prompt. If your token usage doubled, it could be related to: - Conversation context being carried forward across turns - Tool call invocations and results being included in the context - Longer prompts or multi-turn conversations For detailed analytics about your specific token usage patterns and what's being carried forward in each request, I'd recommend checking your call logs in the dashboard or reaching out to Retell AI support for a detailed breakdown of your account's token consumption.

​ To learn more, see the sources I considered: [[1] Platform Changelogs | Retell AI](https://www.retellai.com/changelog) ​

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

### Answer "Conversation context being carried forward across turns" refers to how Retell AI's LLM maintains memory of the entire conversation history when generating responses. What Gets Carried Forward: When the LLM processes each new user input, it includes the previous conversation exchanges (user messages and assistant responses) plus any tool call invocations and their results in its context. [(1) ]() This allows the agent to understand references to earlier parts of the conversation and provide contextually relevant responses. Context Persistence in Retell AI: Retell AI's transformer architecture maintains context throughout the entire conversation, enabling the system to: - **Track conversation history**: Understanding how previous exchanges inform current intent - **Handle clarifications**: Responding appropriately when customers provide additional information - **Manage topic transitions**: Smoothly navigating when conversations shift between different subjects - **Maintain personalization**: Leveraging earlier conversation elements to provide tailored responses Token Impact: Each turn adds tokens to the context window. The conversation history (all prior user/assistant messages) , tool calls, tool results, and your prompts all consume tokens. As conversations get longer with more turns and tool interactions, the total token count sent to the LLM with each request increases, which directly impacts your token usage and costs. This is why longer, multi-turn conversations with multiple tool calls can result in significantly higher token consumption compared to shorter interactions.

​ To learn more, see the sources I considered: [[1] ](https://www.retellai.com/resources/retell-ai-receptionist-understands-calls-better) ​

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

### Answer Agent slowness typically comes from high end-to-end latency. Here's how to diagnose and fix it: Check Your Actual Latency: Monitor latency through the Call History dashboard or use the Get Call API to retrieve detailed breakdowns. [(1) ]() Key metrics include p90 (90% of calls below this value) , median (50th percentile) , and min/max response times. End-to-end latency measures the total time from when the user stops speaking until the AI agent begins responding, including processing time, network delays, and model inference time. Common Latency Bottlenecks: The latency breakdown shows which component is causing delays: - LLM inference - Switch to a faster model if this is the bottleneck [(2) ]() - *TTS (text-to-speech) * - Choose a lower-latency voice provider - Tool calls - Optimize tool endpoints or reduce response size - *ASR (speech recognition) * - Transcription latency between audio chunks and transcribed output [(1) ]() High Interruption Count: If users frequently interrupt, the issue depends on your latency: [(2) ]() - **High latency (e2e p50 > 2.5s) **: Fix latency first using the steps above - **Normal latency**: Decrease agent responsiveness or increase interruption sensitivity Recent Performance Issues: Some users have reported latency increases from ~1500ms to ~3500ms. [(3) ]() One user noted that including knowledge base content directly in the base prompt (as a workaround for another issue) may have contributed to slower performance due to larger prompt sizes. Romanian Language Specific: For Romanian

cold calling, users recommend Deepgram for low latency or OpenAI Whisper for higher accuracy, along with tuning turn-detection, barge-in, and streaming settings to avoid 2-3 second dead air. [(4) ]() Check your call analytics dashboard for latency breakdowns to identify which component needs optimization.

​ To learn more, see the sources I considered: [[1] Check actual latency](https://docs.retellai.com/reliability/check-actual-latency) [[2] Address Metric Issues](https://docs.retellai.com/ai-qa/address-metric-issues) [[3] End to end latency went up from ~1500ms to ~3500ms in the last week](https://discord.com/channels/1186424376300679209/1352660400751116360) [[4] Discord message from @omar.13. on 2026-02-16](https://discord.com/channels/1186424376300679209/1186462578310910112/1473011091792466001) ​

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

### Answer I don't have specific information in the available sources about TTS pronunciation issues with "a.m." and "p.m." abbreviations. However, this type of pronunciation issue typically stems from how the Text-to-Speech (TTS) engine interprets abbreviations. [(1) ]() The TTS system converts text responses into natural-sounding speech, and abbreviations like "a.m." and "p.m." can sometimes be mispronounced depending on the voice model and TTS provider settings. Potential Solutions: Making your agent sound more natural often requires better prompt design, pacing adjustments, and choosing the right voice/TTS settings. [(2) ]() You can try: - Spelling out the time format in your prompts (e.g., "10 AM" instead of "10 a.m.") - Using full words like "in the morning" or "in the evening" instead of abbreviations - Adjusting voice model settings including temperature, speed, and volume controls [(1) ]() - Testing different TTS providers to find one that handles abbreviations better - Using pronunciation hints in your agent's responses if your TTS provider supports SSML or phonetic markup For more detailed TTS configuration and voice quality improvements, I'd recommend checking your voice model settings in the dashboard or reaching out to Retell AI support for provider-specific pronunciation guidance.

​ To learn more, see the sources I considered: [[1] Speech Processing](https://www.retellai.com/glossary/speech-processing) [[2] Discord message from @bruces000 on 2026-05-04](https://discord.com/channels/1186424376300679209/1186462578310910112/1500886739122651266) ​

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

### Answer I don't have specific information in the available sources about a model called "GPT 4.1 fast" or the differences between "GPT 4.1" and "GPT 4.1 fast." The available model options in Retell AI's LLM configuration include: [(1) ]() [(2) ]() [(3) ]() [(4) ]() -

gpt-4.1

-

gpt-4.1-mini

-

gpt-4.1-nano

-

gpt-5

-

gpt-5-mini

-

gpt-5-nano

-

gpt-5.1

-

gpt-5.2

-

gpt-5.4

-

gpt-5.4-mini

-

gpt-5.4-nano

There's no "GPT 4.1 fast" listed in the official model options. If you're experiencing latency issues, you might want to consider: [(2) ]() - Switching to a smaller model like

gpt-4.1-mini

or

gpt-4.1-nano

for faster inference - Using

model_high_priority

set to

true

for more dedicated resources and lower latency (though this comes with higher costs) [(1) ]() If you're looking to optimize for speed, the smaller variants (

gpt-4.1-mini

and

gpt-4.1-nano

) are designed to be faster than the base

gpt-4.1

model while maintaining reasonable performance.

​ To learn more, see the sources I considered: [[1] Update Retell LLM](https://docs.retellai.com/api-references/update-retell-llm) [[2] Get Retell LLM](https://docs.retellai.com/api-references/get-retell-llm) [[3] List Retell LLMs](https://docs.retellai.com/api-references/list-retell-llms) [[4] Create Retell LLM](https://docs.retellai.com/api-references/create-retell-llm) ​

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