Hi, assume I have a knowledge base attached to a query-tool. Question: Do the voice replacements (if any) are applied to the result from query-tool? Or query-tool response are "spoken" exactly as in the knowledge base content ?

its not exactly spoken, the response comes to agent/LLM and u hear that.

The voice replacements would be applied to the result from the query tool, so the response you hear might not be exactly as it is in the knowledge base.

ty @Subhankar Ghosh and @Shaunak another question regarding the query tool. so we use vapi's query-tool with gemini-2.5-flash model. when observing past calls that has query-tool invocation; I've noticed a pattern: - There is data X in knowledgebase. Assume its a simple line of text like "**Object weight: 100kg**" - query-tool is called - query-tool's result is a generated response out of knowledgebase data. It's like "**The object you are interested in weighs 100kg**". - our assistant re-processes the query-tool's response and generates Response B like "**The object you are interested in weighs 100kg, quite heavy isn't it?**" so there is Data X in knowledgebase query-tool generates Response A out of Data X our assistant generates Response B out of Response A is this the expected behaviour? either query-tool's extra response generation or our assistant's extra response generation "feels" like an unecessary operation

Yes, you are right. The double-processing is expected and by design.

and we are not able to see which input was used to call the query-tool right? there is only knowledgebase name in the tool arguments. so we dont "know" which query (as in vector search input vector term) was used to retrieve chunks using the knowledgebase?

Yes