https://retellai.com/?utm_source=discord&utm_medium=seo logo
Powered by
What are common reasons for MCP node not working
# ❓・ai-evy-support
d
@AI Evy What are common reasons for MCP node not working
a
Thanks for tagging me @Dan@beHuman. I'll look through relevant sources and get back to you shortly.
### Answer Based on the sources, here are common reasons MCP nodes may not work properly: 1. Response Variable Extraction Issues - MCP tools may return stringified JSON responses that don't get parsed correctly. [(1) ]() - A regular custom function returns: 
{"customer_id":"6229"}
- Whereas MCP may return: 
"{\"customer_id\": 12345}"
(stringified) - This can cause response variables to not be extracted and transitions to fail. 2. Configuration Not Persisting - In Conversation Flow agents, response variables set in MCP tool blocks may not persist when reopening the block—it reverts to tool selection instead of showing saved variables. [(2) ]() - This appears to be a known bug in Conversation Flow (works correctly in Multi-Prompt type) . 3. Duplicate Functions or Transitions - Errors can occur if there are duplications in functions or transitions in your flow. 4. Tools List Loading Forever - When connecting a custom MCP server, the tools list in the Retell dashboard may get stuck on "Loading..." even if 
/mcp/manifest.json
and 
/mcp/tools
endpoints return correct data via curl. [(3) ]() 5. Server/Protocol Issues - Problems may be related to how the MCP server returns responses (e.g., Make.com using 
/stream
protocol) . [(1) ]() If you're experiencing issues, try refreshing the page or reopening the flow, and check for any duplicate configurations. [(2) ]()
​ To learn more, see the sources I considered: [[1] mcp tool variables not extracted](https://discord.com/channels/1186424376300679209/1426340769924382791) [[2] MCP Tool Configuration Fails to Persist Response Variables](https://discord.com/channels/1186424376300679209/1410233678155616387) [[3] Discord message from @santelo77 on 2025-08-23](https://discord.com/channels/1186424376300679209/1294891368505937952/1408815780912697536) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
Can you elaborate on this "Update prompt for function It’s best to include in the prompt explicitly when is the best time to invoke the mcp tool. For example:" does the prompt go on the mcp node or the transition of another node
@AI Evy Can you elaborate on this "Update prompt for function It’s best to include in the prompt explicitly when is the best time to invoke the mcp tool. For example:" does the prompt go on the mcp node or the transition of another node
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer Based on the available documentation, explicit tool invocation instructions should be included in your prompts (not on transitions) . For Conversation Flow Agents, the Prompt Engineering Guide indicates you should use these principles "within individual node instructions." [(1) ]() The recommended approach is to define clear triggers in your prompt that specify exactly when to call tools. Here's an example structure:
Copy code
///## Tool Usage Instructions
1. Gather initial information about the customer's issue.
2. Determine the type of request:
   - If customer mentions "refund" or "money back":
     → Call function `transfer_to_support` immediately
   - If customer needs order status:
     → Call function `check_order_status` with order_id
Key best practices for tool instructions: - **Use trigger words**: List specific words/phrases that should trigger tool calls - **Define sequences**: Specify when tools should be called in order - **Set boundaries**: Clarify when NOT to call certain tools So for MCP tools, you would include instructions in the node's instruction prompt that explicitly state when the MCP tool should be invoked (e.g., "When the user asks about X, call the MCP tool `tool_name`") . The documentation doesn't provide specific MCP node configuration details beyond what's shown in the API references. If you need more specific guidance on MCP node setup, I'd recommend checking: - [Prompt Engineering Guide]() - [Create Conversation Flow]()
​ To learn more, see the sources I considered: [[1] Prompt Engineering Guide](https://docs.retellai.com/build/prompt-engineering-guide) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
I am getting this error from MCP node when I make a transition to it from a conversation node: "MCP error -32601: Method not found: tools/call"
@AI Evy I am getting this error from MCP node when I make a transition to it from a conversation node: "MCP error -32601: Method not found: tools/call"
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer ### Answer The error 
MCP error -32601: Method not found: tools/call
indicates that your MCP server doesn't support the 
tools/call
method that Retell AI is trying to invoke. [(1) ]() This is typically an MCP server configuration issue, not a Retell AI issue. The error means: 1. Your MCP server endpoint may not implement the required method — Retell AI's MCP node expects the server to handle 
tools/call
requests, but your server isn't recognizing that method. 2. Protocol mismatch — Some MCP servers (like Make.com using 
/stream
protocol) may return responses in formats that don't align with what Retell expects. [(2) ]() Things to check: - Verify your MCP server correctly implements the 
tools/call
endpoint/method [(1) ]() - Ensure your MCP Server Endpoint URL is correct and the server is running [(3) ]() - Check that the tool you selected in the MCP node actually exists on your server - Review your MCP server logs to see how it's receiving and processing the request If using a third-party MCP service (Zapier, Make.com, custom API) , confirm their documentation shows support for the standard MCP 
tools/call
method. [(1) ]() One user with similar MCP issues ended up switching to custom functions instead, which worked well for their use case. [(4) ]()
​ To learn more, see the sources I considered: [[1] Connect any AI Voice Agent to MCP with Retell AI MCP Node (Guide) | Retell AI](https://www.retellai.com/blog/connect-any-ai-voice-agent-to-mcp-with-retell-ai-mcp-node) [[2] mcp tool variables not extracted](https://discord.com/channels/1186424376300679209/1426340769924382791) [[3] MCP Node](https://docs.retellai.com/build/conversation-flow/mcp-node) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy I am curious to know how things on retell are injected into context. Specifically, i want to know how fine tune examples are injected into context. I am worried that if I have a conversation block with a big prompt and then a lot of fine-tuning examples that this will cloud the context too much and not make the LLM operate as desired
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer --- The available documentation doesn't provide specific technical details about how fine-tune examples are injected into context or the exact mechanics of context assembly in Retell AI. What I can share from the sources: Fine-tuning examples are available at the node level: - Both the Extract Dynamic Variable Node and Function Node have a Fine-tuning Examples setting that "can finetune transition." [(1) ]() [(2) ]() General context considerations: - There is a prompt token limit of approximately 32,768 tokens. [(3) ]() - Retell's engineers recommend keeping prompt size below 8,000 tokens for optimal performance, noting that larger prompts can "significantly reduce stability, causing the LLM to become confused and respond inaccurately" and also increase latency and costs. Conversation Flow's purpose: - The Conversation Flow feature uses a node-based structure specifically designed to help "mitigate AI hallucinations" by providing structured pathways that guide responses based on user input. [(4) ]() - This approach breaks conversations into distinct nodes with their own instructions, which may help manage context more effectively than one large prompt. For the specific technical details on how fine-tune examples are weighted or injected relative to node prompts, I don't have that information in the available sources. You may want to reach out via [Contact Sales]() or check the [GitHub]() for more technical implementation details. Would you like me to try to answer this more generally based on my
​ To learn more, see the sources I considered: [[1] Extract Dynamic Variable Node](https://docs.retellai.com/build/conversation-flow/extract-dv-node) [[2] Function Node Overview](https://docs.retellai.com/build/conversation-flow/function-node) [[3] Prompt exceeds the limit of 32768 tokens, please contact support if you need longer context.](https://discord.com/channels/1186424376300679209/1432544872421789789) [[4] Training & Customizing AI Voice Agents | Retell AI](https://www.retellai.com/blog/training-and-customizing-voice-agents-with-retell-ai) ​
knowledge?
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy Can batch tests be run mutiple times without me having to click run everytime?
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer --- Yes, you can run batch tests programmatically via the API without manually clicking each time. [(1) ]() Using the Create Batch Test API endpoint, you can automate batch test execution:
Copy code
javascript
import Retell from 'retell-sdk';
const client = new Retell({
 apiKey: 'YOUR_RETELL_API_KEY',
}) ;
const batchTestResponse = await client.tests.createBatchTest({
 response_engine: { llm_id: 'llm_id', type: 'retell-llm' },
 test_case_definition_ids: ['string'],
}) ;
console.log(batchTestResponse.test_case_batch_job_id) ;
This allows you to: - Trigger batch tests from scripts or CI/CD pipelines - Pass an array of test case definition IDs (1-200 elements per batch) - Optionally reserve concurrency for batch processing The response includes a 
test_case_batch_job_id
and status (
in_progress
or 
complete
) along with pass/fail/error counts. For scheduling or repeated automated runs, you'd need to build a wrapper (e.g., cron job, scheduled task) that calls this API endpoint at your desired intervals—the documentation doesn't show a built-in scheduling feature for batch tests specifically.
​ To learn more, see the sources I considered: [[1] Create Batch Test](https://docs.retellai.com/api-references/create-batch-test) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
d
@AI Evy I am struggling with my agent a lot right now. In testing where i am isolating nodes and makign sure they work as needed, I am getting great results, but when I am tesing the whole flow via voice for some reason the last node just breaks down and works very poorly ( I have a 2 node system setup). I am assuming this is due to context polution. What can I do?
a
### Answer ### Answer Based on the sources, here are some approaches for when nodes work in isolation but break down in the full flow: 1. Consider Breaking Down the Problem Node The documentation recommends breaking down a node when: - The node handles multiple complex logic paths - The LLM struggles with consistency (hallucinations or incorrect responses) - You need different settings (model, temperature) for different parts - The conversation flow becomes hard to follow or debug [(1) ]() "Breaking complex nodes into smaller, focused nodes often improves reliability." 2. Voice vs Chat Differences One user reported their conversation flow worked in chat mode but not in voice mode. The recommendation was to "Check Voice-Specific Settings." [(2) ]() 3. Known Issue with Function Nodes + Response Timing There was a resolved incident where conversation flow agents would stay silent after a function node. This happened "when a function and response is happening at same time" due to a message order issue. [(3) ]() 4. Structural Approaches - Try using Components to encapsulate and isolate logic—this can help "hide detailed logic inside a focused sub-flow" and keep the canvas cleaner. [(4) ]() - Ensure your transition conditions between nodes are clear and not conflicting. - Each node can be "fine-tuned independently for better performance." [(1) ]() If the second node specifically is the issue, try simplifying its instructions or splitting it further. The node-based structure is designed to provide "preci
​ To learn more, see the sources I considered: [[1] Node Overview](https://docs.retellai.com/build/conversation-flow/node) [[2] Conversation flow working in test via chat but not in voice](https://discord.com/channels/1186424376300679209/1357369181363769566) [[3] Conversation Flow agent potentially stay silent after function node](https://retell.discourse.group/t/conversation-flow-agent-potentially-stay-silent-after-function-node/135) [[4] Components](https://docs.retellai.com/build/conversation-flow/components) ​
se control" and make "easier debugging" by isolating issues in specific conversation paths.
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
Is there no way to clear all context when entering a node? The node is clearly able to work fine when tested in isolation so the tasks are not too complex
@AI Evy Is there no way to clear all context when entering a node? The node is clearly able to work fine when tested in isolation so the tasks are not too complex
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer Based on the available sources, there is no documented feature to clear or reset context when entering a node in Retell AI Conversation Flow. However, here are some alternative approaches that may help: 1. Use Fine-tune Examples You can supply fine-tune examples to guide the LLM's behavior at specific nodes. You provide a transcript as context and the expected response or transition result—this can help "finetune agent's response" and transitions without needing to clear prior context. [(1) ]() *2. Consider Flex Mode (with caution) * Flex Mode compiles your conversation flow into tasks and allows the agent to navigate dynamically. However, there are trade-offs: - Node-level Knowledge Base is ignored in Flex Mode [(2) ]() - Performance may degrade with more than 20 nodes - "LLM might not always follow the static text instruction" 3. Structural Redesign Since "each node can be optimized independently for better performance," you might try: [(3) ]() - Using different model settings (model, temperature) for the problematic node - Breaking the node into smaller, more focused nodes - Using node-specific Knowledge Base to limit what information is available at that point [(4) ]() The Conversation Flow's node-based structure is specifically designed to provide "precise control" and reduce context drift through "transition conditions" that define when and where to move next. [(3) ]() If context pollution is the core issue, I don't have documentation showing a direct "clear context" option.
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy I want to learn how retell is built to better understand how I should use it. Are there any resources that explain how retell works?
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer ### Answer Here's what the sources explain about how Retell AI's Conversation Flow works: How Context is Assembled Based on community discussion, the LLM sees context structured roughly like this: [(1) ]() - System prompt - General instruction (global prompt) - Instruction from the current node - RAG chunks (Knowledge Base) Core Components The Conversation Flow system consists of: [(2) ]() - **Global Settings**: Configuration that applies to the entire conversation (global prompt, voice, language settings, agent-wide parameters) - **Nodes**: The basic unit of conversation flow—conversation nodes for dialogue, function nodes for API calls, logic nodes for branching, end nodes for termination - **Edges**: Connections between nodes that define transition logic (condition-based transitions, default fallback paths, dynamic routing) - **Functions**: Reusable capabilities invoked from function nodes How It Works Every node defines a small set of logic, and transition conditions determine which node to transition to. The global prompt is available in every node and influences all response generation. [(3) ]() Dialogue Management Retell's dialogue management includes state tracking (what's been said/confirmed) , decision logic, error handling, context awareness across turns, and goal-oriented flow. [(4) ]() Resources to Learn More - [Conversation Flow Overview]() [(2) ]() - [Node Overview]() [(5) ](https:/
/docs.retellai.com/build/conversation-flow/node) - [Retell AI's Advanced Conversation Flow blog post]() [(6) ]() - [Platform Changelogs]() [(5) ]()
​ To learn more, see the sources I considered: [[1] Discord message from @sikntwisted_ on 2025-03-07](https://discord.com/channels/1186424376300679209/1294891368505937952/1347470163297046561) [[2] Conversation Flow Overview](https://docs.retellai.com/build/conversation-flow/overview) [[3] Step 1: Configure global settings](https://docs.retellai.com/build/conversation-flow/global-setting) [[4] Dialogue Management](https://www.retellai.com/glossary/dialogue-management) [[5] Node Overview](https://docs.retellai.com/build/conversation-flow/node) [[6] Retell AI’s Advanced Conversation Flow | Retell AI](https://www.retellai.com/blog/unlocking-complex-interactions-with-retell-ais-conversation-flow) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy You say that's it's structured through a system prompt at the first stage, could you tell me more about that? Are there any resources on how context is injected into retell?
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer ### Answer The available documentation doesn't provide detailed technical specifications on exactly how context is assembled and injected into the LLM at runtime. What the sources do explain: For *Multi-Prompt Agents (LLM States) *, the system prompt is constructed as: [(1) ]() - 
System prompt (with state) = general prompt + state prompt
- 
System prompt (no state) = general prompt
Similarly, tools are assembled as: - 
Tools of LLM (with state) = general tools + state tools + state transitions
- 
Tools of LLM (no state) = general tools
From community discussion, one user's understanding (which was confirmed as "more or less correct") suggests the context structure looks like: [(2) ]() - System prompt - General instruction - Instruction from the current node - RAG chunks For Conversation Flow, the documentation explains that: [(3) ]() - Global Settings apply to the entire conversation (default voice, language preferences, global prompt) - Nodes dictate specific interactions with their own prompts - The global prompt "sets the context for conversations across all nodes" to maintain consistency Regarding prompt engineering, the Prompt Engineering Guide recommends using sectional prompts with sections like Identity, Style Guardrails, Response Guidelines, Task Instructions, and Objection Handling. [(4) ]() Unfortunately, there isn't detailed documentation on the exact injection mechanics or context window management. You may want to reach out via [Contact Sales]() or check [GitHub](
com/RetellAI/) for more technical details.
​ To learn more, see the sources I considered: [[1] Update Retell LLM](https://docs.retellai.com/api-references/update-retell-llm) [[2] Discord message from @sikntwisted_ on 2025-03-07](https://discord.com/channels/1186424376300679209/1294891368505937952/1347470163297046561) [[3] Retell AI’s Advanced Conversation Flow | Retell AI](https://www.retellai.com/blog/unlocking-complex-interactions-with-retell-ais-conversation-flow) [[4] Prompt Engineering Guide](https://docs.retellai.com/build/prompt-engineering-guide) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy why do my dynamic variables in tests (Test Variables & Mocks) not save?
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer ### Answer This is expected behavior. According to Retell AI support, dynamic variables set in the test agent are runtime-only and are not saved to the database. [(1) ]() From a support response: > "That is the expected behavior. It's a runtime test. All input will be used only in runtime. Nothing will be saved to db" Workaround options: 1. Use Default Values in Agent Settings — You can set fallback/default values for dynamic variables in the agent's settings panel. These will persist and be used when variables aren't provided at runtime. [(2) ]() 2. Access fallback settings — Navigate to the agent settings where you can define default values for your dynamic variables. This is a known pain point—other users have reported the same frustration when testing with many variables. [(1) ]()
​ To learn more, see the sources I considered: [[1] Dynamic variables not persisted for test agent](https://discord.com/channels/1186424376300679209/1419312690110660740) [[2] Dynamic variables in test agent not working anymore](https://discord.com/channels/1186424376300679209/1415366568812220577) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy There was an email not that long ago that retell sent out to do with SIP trunking and making a change to it if it's being used. Can you tell me more about that?
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer Based on a Discord message, there was indeed an email from Retell regarding a telephony network update for SIP trunking users. Here's what was communicated: [(1) ]() If you use IP ACL for SIP Authentication: - Keep all previously allowlisted Retell AI IP ranges in place - Add the new CIDR to your allowlist: 
18.98.16.120/30
- Ensure your network firewalls allow this CIDR - Confirm your SIP trunking provider has IP-based authentication configured correctly - Some providers (e.g., Telnyx) may require additional settings If you use Credentials for SIP Authentication: - Ensure your firewalls allow: 
18.98.16.120/30
- No further authentication changes are required Important note: If you purchased your number directly through Retell (not via SIP trunking) , you don't need to take any action—this only applies to users with custom SIP trunk configurations. The IP block 
18.98.16.120/30
is also documented as the traffic range for all regions that you can whitelist for Retell's SIP server. [(2) ]()
​ To learn more, see the sources I considered: [[1] Discord message from @briancorbett on 2025-12-11](https://discord.com/channels/1186424376300679209/1294891368505937952/1448753936008740965) [[2] Custom Telephony Overview](https://docs.retellai.com/deploy/custom-telephony) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy Is there any limit or recommended limit to the amount of characters, lines or sentences that should be inside a conversation node
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
​ To learn more, see the sources I considered: [[1] Understand concurrency & limits](https://docs.retellai.com/deploy/concurrency) [[2] Prompt exceeds the limit of 32768 tokens, please contact support if you need longer context.](https://discord.com/channels/1186424376300679209/1432544872421789789) [[3] Master The Future of Voice AI - Conversation Design & Agent Persona](

https://www.youtube.com/watch?v=42V0bDQIsb0▾

) [[4] Node Overview](https://docs.retellai.com/build/conversation-flow/node) ​
### Answer The documentation doesn't specify a character, line, or sentence limit for individual conversation nodes. However, there are relevant overall limits and recommendations: Hard Limit: - The maximum prompt token length when using Retell LLM is 32,768 tokens by default. [(1) ]() - Prompts over 3,500 tokens will be charged extra. Recommended Limits (from Retell engineers) : - Retell engineers recommend keeping prompt size below 8,000 tokens for optimal performance. [(2) ]() - Using over 33,000 tokens "can significantly reduce stability, causing the LLM to become confused and respond inaccurately." - Larger prompts also increase latency and costs. Community Recommendation: - One user shared they keep prompts below 1,000 tokens because "after that it seems to hallucinate more." [(3) ](

▾

) Best Practice: - The node-based structure is designed so you can break complex workflows into smaller, focused nodes—each with its own logic and behavior—rather than having one large node. [(4) ]() - "Breaking complex nodes into smaller, focused nodes often improves reliability." If you're experiencing issues with a node, consider splitting it into multiple smaller nodes rather than keeping everything in one large prompt.
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy How do I know how many tokens a prompt uses
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer You can check token usage in the call details after a call completes. [(1) ]() Keep in mind that the token count includes more than just your prompt—it also factors in the transcript between agent and user, plus tool call history and results. External Token Counting Tool: One community member shared a helpful approach: use a token counter tool to measure your prompt before deploying. In a workshop, they mentioned: "I use this thing I've just posted in the chat to measure my prompt so it's below 1000 tokens." [(2) ](

▾

) Key thresholds to know: - Prompts up to 3,500 tokens follow the base rate [(3) ]() - Prompts exceeding 3,500 tokens are charged proportionally [(4) ]() - Maximum prompt token length is 32,768 tokens [(3) ]() For pre-deployment estimation, you can use external tokenizer tools (like OpenAI's tiktoken or online token counters) to get an approximate count before testing.
​ To learn more, see the sources I considered: [[1] Token Surcharge when prompt tokens is under 3500](https://discord.com/channels/1186424376300679209/1442646663335510179) [[2] Master The Future of Voice AI - Conversation Design & Agent Persona](

https://www.youtube.com/watch?v=42V0bDQIsb0▾

) [[3] Understand concurrency & limits](https://docs.retellai.com/deploy/concurrency) [[4] 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.
d
@AI Evy I can't find the token usage per prompt in teh call details/
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer The call response object includes an 
llm_token_usage
field that provides token usage data. [(1) ]() Here's what's available:
Copy code
javascript
"llm_token_usage": {
  "values": [
    123
  ],
  "average": 123,
  "num_requests": 123
}
This field is: - Available after the call ends - Not populated if using custom LLM, realtime API, or if no LLM call is made You can retrieve this via the Get Call API endpoint or the List Calls endpoint. , [(2) ]() However, this shows aggregate token usage across the call—not a per-prompt breakdown showing exactly how many tokens each individual prompt consumed. The documentation doesn't indicate a way to see token counts for each specific prompt or node within a call. For pre-deployment estimation of your prompt's token count, you'd need to use external tokenizer tools (like OpenAI's tiktoken) as mentioned in the previous discussion.
​ To learn more, see the sources I considered: [[1] List Calls](https://docs.retellai.com/api-references/list-calls) [[2] Get Call](https://docs.retellai.com/api-references/get-call) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy does having unused nodes in the agent pollute context?
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer Based on the available sources, the documentation doesn't explicitly address whether unused nodes pollute context. However, based on how Conversation Flow works, the LLM context appears to include: [(1) ]() - System prompt - General instruction (global prompt) - Instruction from the current node - RAG chunks This suggests that only the current node's instruction is injected into context at any given time—not all nodes in the agent. [(2) ]() The system is designed so that "every node defines a small set of logic, and the transition condition is used to determine which node to transition to." Since nodes are visited based on transition conditions and edges, unused nodes that are never transitioned to shouldn't be included in the LLM context during a call. [(3) ]() That said, I don't have explicit documentation confirming this behavior. If you're concerned about context pollution, you could: - Remove truly unused nodes from your flow - Use Components to encapsulate and organize logic into focused sub-flows [(4) ]() If you need definitive confirmation, you may want to reach out via [Contact Sales]() or check [GitHub]() .
​ To learn more, see the sources I considered: [[1] Discord message from @sikntwisted_ on 2025-03-07](https://discord.com/channels/1186424376300679209/1294891368505937952/1347470163297046561) [[2] Conversation Flow Overview](https://docs.retellai.com/build/conversation-flow/overview) [[3] Node Overview](https://docs.retellai.com/build/conversation-flow/node) [[4] Components](https://docs.retellai.com/build/conversation-flow/components) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy Are there any more uses to components ohter than grouping logic together so that it can be reused? Do components affect context in any way?
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer Based on the documentation, Components offer several benefits beyond just reusing logic: Key Benefits of Components: [(1) ]() - **Reuse**: Build once, drop into many agents and flows - **Consistency**: Keep behavior uniform across use cases (e.g., identity check) - **Clean canvas**: Hide detailed logic inside a focused sub-flow - **Faster iteration**: Update a shared Component to improve every agent that uses it Regarding Context: The documentation doesn't explicitly state that Components affect context differently than regular nodes. The global prompt of the main Conversation Flow is applied to all component nodes implicitly. When you use a Component, it functions as a mini flow—the nodes inside behave the same as nodes in your main flow. Tools defined within the component "behave the same as in the main flow." Version Control for Production: One notable feature is how Components behave at publish time: - When you publish an agent, the published version snapshots shared Components as local copies - This prevents future library updates from changing already-published calls - Your draft stays linked to the shared Component, so you continue to benefit from library updates while iterating This provides stability for production while allowing ongoing development. The sources don't document any special context isolation or clearing behavior specific to Components—they appear to inherit and work within the same context model as regular nodes.
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] Components](https://docs.retellai.com/build/conversation-flow/components) ​
d
@AI Evy When I am using the chat interface to test my agent, I often click on debug and then regenrate 10 answers so that I can test how my agent answers to certain things faster. Sometimes however I don't get 10 answers generated, but some number less than 10. Why does this happen?
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer The available documentation doesn't explicitly explain why fewer than 10 answers are sometimes generated when using "Regenerate 10 Times." [(1) ]() The documentation describes the feature as a way to "test response consistency" by clicking "Regenerate 10 Times" to check variations of an answer. [(2) ](

▾

) In the video walkthrough, the feature is shown generating responses to "check the variations of this answer" and verify the agent "will not go off the reel." However, the sources don't document specific reasons why you might receive fewer than 10 responses. Possible causes could include timeouts, rate limiting, or errors during generation—but this isn't confirmed in the available documentation. If this is happening consistently, you may want to reach out via [Contact Sales]() or check [GitHub]() for more details.
​ To learn more, see the sources I considered: [[1] Debug your agent response](https://docs.retellai.com/test/llm-playground-debug) [[2] CF Testing Suite | Retell AI Feature](

https://www.youtube.com/watch?v=inMXRs9dbT4▾

) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy We have an existing phone number that we want to connect to retell. The number is a vanity number from RingCentral. Are there any instructions for connecting via SIP trunking for this case?
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer --- Retell AI doesn't have specific documentation for RingCentral SIP trunking, but you can use the general custom telephony integration approach. There are two methods available: [(1) ]() *Method 1: Elastic SIP Trunking (Recommended) * If RingCentral supports elastic SIP trunking, you would: 1. Set up a SIP trunk in RingCentral and configure your vanity number to point to it 2. Import that number to Retell Retell provides detailed guides for Twilio, Telnyx, and Vonage—you can use these as a reference for RingCentral since "other telephony providers that support SIP trunks are also supported." Method 2: Dial to SIP URI If RingCentral doesn't support elastic SIP trunking or you have a more complex setup: - Retell SIP server URI: 
sip:sip.retellai.com
- IP block to whitelist: 
18.98.16.120/30
(All regions) - Supported transport methods: TCP (recommended) , UDP, TLS For inbound calls with different transport: - TCP: 
sip:sip.retellai.com;transport=tcp
- UDP: 
sip:sip.retellai.com;transport=udp
- TLS: 
sip:sip.retellai.com;transport=tls
With this method, you'd call the [Register Phone Call API]() to get a 
call_id
, then dial to: 
sip:{call_id}@sip.retellai.com
Note: Retell won't throw errors if the configuration is incorrect—you'll only know if the setup works once a call is made. If calls aren't connecting, check your SIP trunk logs with RingCentral.
​ To learn more, see the sources I considered: [[1] Custom Telephony Overview](https://docs.retellai.com/deploy/custom-telephony) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy Can you see if there have been any messages to do with RingCentral and connecting it to retell or other mentions of it in this discord
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer Based on the sources I have access to, I don't see any Discord messages specifically mentioning RingCentral or connecting it to Retell. The Discord messages I can search include discussions about custom SIP trunking with providers like Twilio, Telnyx, and custom B2BUA setups, but no specific RingCentral references. [(1) ]() One user mentioned having "a custom sip trunk that is used as a universal solution to adapt to different vendor trunks" and using jambonz in some places, but RingCentral wasn't specifically discussed. For your RingCentral integration, the general custom telephony guidance would apply—you'd need to determine if RingCentral supports elastic SIP trunking or if you'd need to use the dial-to-SIP-URI method. [(2) ]() You might want to: - Search the Discord server directly using Discord's search function for "RingCentral" - Post a new question in the community asking if anyone has successfully connected RingCentral - Check with RingCentral support about their SIP trunking capabilities
​ To learn more, see the sources I considered: [[1] Discord message from @michaelb_30512 on 2025-09-23](https://discord.com/channels/1186424376300679209/1186462578310910112/1420067324408369184) [[2] Custom Telephony Overview](https://docs.retellai.com/deploy/custom-telephony) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy How do you make a variable for the current time EST on retell
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer You can use Retell's built-in system variable with a timezone suffix to get the current time in EST (Eastern Standard Time) . [(1) ]() Use this syntax in your prompt or configuration:
Copy code
{{current_time_America/New_York}}
This will return the current time formatted like: "Thursday, March 28, 2024 at 11:46:04 PM EDT" How it works: - 
{{current_time}}
defaults to 
America/Los_Angeles
timezone - 
{{current_time_[timezone]}}
lets you specify any timezone - For EST/EDT, use 
America/New_York
as the timezone identifier , [(2) ]() This is a default system variable—no configuration required. [(1) ]()
​ To learn more, see the sources I considered: [[1] Dynamic Variables](https://docs.retellai.com/build/dynamic-variables) [[2] 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.
d
@AI Evy Is there a notification that you can setup to alert you anytime a call is made to an agent?
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
### Answer Yes, you can set up webhooks to receive real-time notifications when calls are made. [(1) ]() Retell AI supports the following webhook events: - 
call_started
 – Triggered when a new call begins (provides basic call information) - 
call_ended
 – Triggered when a call completes, transfers, or encounters an error - 
call_analyzed
 – Triggered when call analysis is complete To set this up: You configure a webhook URL in your agent settings, and Retell will 
POST
the payload to your endpoint when events occur. Here's a sample payload for a 
call_started
or 
call_ended
event:
Copy code
json
{
  "event": "call_ended",
  "call": {
    "call_type": "phone_call",
    "from_number": "+12137771234",
    "to_number": "+12137771235",
    "direction": "inbound",
    "call_id": "Jabr9TXYYJHfvl6Syypi88rdAHYHmcq6",
    "agent_id": "oBeDLoLOeuAbiuaMFXRtDOLriTJ5tSxD",
    "call_status": "registered",
    "start_timestamp": 1714608475945,
    "end_timestamp": 1714608491736,
    "disconnection_reason": "user_hangup"
  }
}
Note: If the call did not connect (like dial failed) , the 
call_started
webhook event will not be triggered—only 
call_ended
and 
call_analyzed
will fire. From there, you can use the webhook to trigger Slack messages, email alerts, or any other notification system you prefer. [(2) ]() , [(3) ]()
​ To learn more, see the sources I considered: [[1] Webhook Overview](https://docs.retellai.com/features/webhook-overview) [[2] Webhook](https://www.retellai.com/glossary/webhook) [[3] Retell AI Webhooks | AI Voice Agents With Live Data | Retell AI](https://www.retellai.com/blog/retell-ai-webhooks-feature) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
d
@AI Evy Is there any way to do a healthcheck or tell that your agent is up and working without calling it?
a
Thanks for the follow-up, @Dan@beHuman. I'll get back to you soon.
​ To learn more, see the sources I considered: [[1] Health Checks](https://discord.com/channels/1186424376300679209/1399426187629039797) [[2] Retell AI Webhooks | AI Voice Agents With Live Data | Retell AI](https://www.retellai.com/blog/retell-ai-webhooks-feature) ​
### Answer Based on the available sources, there isn't a dedicated health check endpoint to verify your agent is operational without making a call. [(1) ]() However, one community member asked a similar question, and a Retell support engineer suggested a workaround: Using Inbound Call Webhook for Immediate Transfer: You can set up an agent with an immediate transfer using the inbound call webhook. When a call hits your number, the webhook can immediately return a transfer destination without engaging the AI. This approach lets you test that your system is connected end-to-end without the AI actually processing a conversation (and incurring charges for AI usage) . For setup details, see the [Inbound Call Webhook documentation]() . Alternative Approaches: You could also: - Use the API to verify connectivity (e.g., calling [Get Agent]() or [List Calls]() endpoints to confirm your API key works and the agent exists) - Set up webhooks to monitor for 
call_started
events to confirm the system is receiving calls [(2) ]() But for a true end-to-end health check without a full AI conversation, the immediate transfer via inbound webhook is the suggested approach. [(1) ]()
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
PreviousNext
Powered by