Two-way Conversation

Two-way conversation agents run real-time AI phone calls. The agent listens, understands, asks follow-up questions, and produces transcripts, summaries, ratings, recordings, and exports after the call.

How two-way calls work

The call connects through Plivo.
STT transcribes the caller in real time.
The LLM decides the next response using the prompt and linked knowledge bases.
TTS speaks the response back to the caller.
The call continues until the script is complete, silence timeout triggers, or total timeout is reached.
After the call, Veytrix stores transcript, summary, rating, recording, and webhook events based on settings.

Best use cases

Use case	What the agent does
Candidate screening	Asks eligibility, experience, skills, salary, location, and availability questions.
Lead qualification	Checks interest, budget, timeline, requirements, and next-step readiness.
Feedback calls	Collects structured feedback and asks follow-up questions when answers are unclear.
Support triage	Understands the issue, asks clarifying questions, and records structured details.

Mandatory tabs

Tab	Required?	What to fill
Agent	Required	Agent type, welcome message, prompt, tone, required questions, and closing behavior.
LLM	Required	Provider, model, credential, max tokens, temperature, and knowledge bases.
Audio	Required	Language, STT provider/model, TTS provider/model, voice, speed, and credentials.
Engine	Required	Pipeline or Gemini Live, interruption, endpointing, and response timing settings.
Call	Required	Plivo account/number, silence timeout, call timeout, recording, voicemail detection, final message.
Analytics	Recommended	Rating template, summary, extraction prompt, and webhook URL.
Inbound	Optional	Assign a phone number if this agent should receive inbound calls.

Agent tab setup

Select Real conversation as the agent type.
Write a short welcome message that explains who is calling and why.
Write the system prompt with objective, question order, follow-up rules, allowed knowledge, and closing message.
Keep responses short so the call feels natural.

You are an AI screening agent for a Java Full Stack role.
Greet the candidate by name and confirm if they can speak now.
Ask about Java, Spring Boot, React, SQL, and project experience.
Ask one follow-up when the answer is vague.
If the candidate asks job details, answer only from the linked knowledge base.
Collect availability, expected salary, and current location.
End politely after all required questions are complete.

LLM tab setup

Provider - choose OpenAI, Gemini, Groq, or Cerebras based on your configured keys.
Model - select the model your workspace uses for real-time calls.
Credential - select the API key from Settings > AI Services.
Max tokens - keep responses controlled; too high can make the agent speak too long.
Temperature - use lower values for consistent screening and higher values only for flexible conversations.
Knowledge bases - attach FAQs, role details, policy notes, or product information.

Audio tab setup

Language - choose the speaking language expected for the call.
STT - choose Sarvam, Deepgram, or Cartesia and select a transcription model.
Keywords - add names, role titles, course names, or technical words that must be recognized correctly.
TTS - choose Cartesia, Sarvam, or ElevenLabs for the agent voice.
Voice and speed - test speed on phone audio; faster voices can reduce clarity.

Engine tab setup

Pipeline - separate STT, LLM, and TTS providers. Best for provider control and knowledge bases.
Gemini Live - speech-to-speech for lower latency when separate provider control is less important.
Interrupt settings - tune how quickly the agent stops speaking when the caller interrupts.
Endpointing - tune how long the system waits before treating caller speech as complete.

Call tab setup

Select the Plivo account or outbound number.
Set Hangup on user silence so silent calls end automatically.
Set Total call timeout to protect usage and avoid very long calls.
Enable voicemail detection when campaign quality depends on live answers.
Enable recording only when your workspace policy allows it and S3 storage is configured.
Write a final call message the agent can say before ending.

Analytics tab setup

Select a rating template so completed calls appear with the correct score columns.
Enable summarization for quick review from Call History.
Use extraction to capture fields such as notice period, salary, city, or interest level.
Add a webhook URL if another system needs call or rating results automatically.

Inbound tab setup

Use this only when callers should call a number and reach this agent. Assign one available number from My Numbers. If the number is already assigned to another agent, remove that assignment first.

Test checklist

Run a test call to your own phone number.
Interrupt the agent once to confirm barge-in behavior.
Answer vaguely and confirm the agent asks a useful follow-up.
Ask a knowledge-base question and confirm it answers from your document.
Check Call History transcript and summary.
Check Ratings to confirm the expected rubric columns are generated.

Two-way conversation calls consume voice minutes for answered talk time. Check usage and remaining minutes before launching large campaigns or batches.