Agent not speaking tool result from Query Tool (Vapi)
Call iD: c6c46f3f-57ae-4341-955f-8be8eeb5abe9
We’re using Vapi’s query tool with (GPT-4o model for assistant)
Problem:
After a successful tool call (result is returned cleanly as a quoted string), the agent remains silent and does not say the result — unless the user repeats the question.
What we tried:
--> Model set to gpt-4o
--> Tool returns string like "result": ""Studios start at AED 460K...""
--> Tool config includes:
{
"type": "query",
"messages": [{ "type": "request-start", "blocking": true }]
}
--> Added instructions in the system prompt:
The agent hears the query, calls the tool correctly, receives the result — but says nothing after tool return. No fallback or recovery logic is triggered.
Would appreciate guidance on whether this is:
A tool response formatting issue?
Something in how GPT-4o handles tool messages?
A known issue in Vapi’s execution pipeline?
Kindly
Suggest a Verified Workaround
We’re using Vapi’s query tool with (GPT-4o model for assistant)
Problem:
After a successful tool call (result is returned cleanly as a quoted string), the agent remains silent and does not say the result — unless the user repeats the question.
What we tried:
--> Model set to gpt-4o
--> Tool returns string like "result": ""Studios start at AED 460K...""
--> Tool config includes:
{
"type": "query",
"messages": [{ "type": "request-start", "blocking": true }]
}
--> Added instructions in the system prompt:
- Always parse the result from query_tool_NEW as a quoted string.
- After the tool result is received, speak it naturally and follow up with a question.
- Also tested with request-complete, async false, and without any tool call blocking.
The agent hears the query, calls the tool correctly, receives the result — but says nothing after tool return. No fallback or recovery logic is triggered.
Would appreciate guidance on whether this is:
A tool response formatting issue?
Something in how GPT-4o handles tool messages?
A known issue in Vapi’s execution pipeline?
Kindly
Suggest a Verified Workaround
