LOCAL WORKSPACE

One conversation.
Every step visible.

Input, knowledge, response aur awaaz ka real-time process.

Services

Conversation

Typed input output

Choose STT, LLM and TTS independently. Enable the other provider as fallback for each stage.

Active languages 0 / 4

Choose 1 to 4 from the 14 options.

Conversation language

Loading language options...

Detected language will appear here.

Starter voice Fall back to Flash-Lite if the selected Live model is unavailable

Gemini Live uses its own voice. Selected STT supplies accepted input. Main and action updates join the same conversation. Switching restarts microphone capture safely.

Backend voice LLM Use other LLM if unavailableVoice output (TTS) Use other TTS if unavailable

Deepgram speech input; Gemini fallback.

ElevenLabs voice Stream input while typing Sales conversation enabled

Starter combines voice response updates into one conversation. Text chat keeps its direct response path. Fallback switches once before output starts; partial answers are not replayed. Text chat keeps its existing Flash flow.

Next voice agentmain

Content RAG

Applies when you connect.

Use your past 15 days of chat in answers. Applies to the next input; an ongoing answer keeps its setting. Saving and 15-day cleanup continue. Current-session context stays available.

Microphone

Voice input (STT) Use STT fallback if unavailable

Stop the microphone before changing STT, LLM or fallback settings.

Noise cleanup

Browser echo cancellation, noise suppression and gain control are always requested. Stop before switching. First usable voice input to LLM target: 600?700 ms. This is a target, not a guaranteed measured delay.

Voice Focus is optional; checking setup...

Microphone off

Streaming transcript counts and measured timings appear here.

Interruption: Waiting for confirmed speech

Sound alone does not stop playback. Confirmed speech interrupts; echo and ordinary haan/hmm are filtered.

LIVE PIPELINE

Ready to connect

Connect a session, then type or use the microphone.

01InputWaiting

02RulesWaiting

03RAGWaiting

04Main LLMWaiting

05RoutingWaiting

06OutputWaiting

User input

No active input

Enter to submit · Shift+Enter for a new line

Live transcript below is provisional; speech checks still decide accepted input.

Starter voice response

Ready for voice input

Checking prepared website knowledge...

Live spoken reply. Playback status appears above.

Conversation output

—

Answer yahan stream hoga.

Early draft

Provisional

Voice: early text goes directly to speech. RAG continues in parallel; any necessary completion adds to the same answer. Chat drafts remain inspection-only.

Waiting for live input…

RAG evidence

Waiting

Waiting for a query…

ACTION SELECTION

Final intent

Current request → action RAG → dedicated selector → selected intent. Questions keep the main response; action-only requests get a short action reply. Global controls use fast matching.

Intent selector LLM Use other selector if unavailable

Submit a request to see the selected intent.

Action retrieval, model attempts and final selection

No selection yet.

Selection preview only. Drafts require owner review. This panel does not execute website actions.

Timing breakdown

0 provider requests

Context size appears with each request.

Server durations share one monotonic turn clock. Browser audio/render timings use their own clock. Audio render is not a physical speaker measurement.

Request process

Time / clock	Stage	What happened

Event details

Select any process row.

Selected rules and connection contract

No profile selected.

Sibling modules are resolved from this checkout. Host owns routing and delivery. RAG remains extractive; no hidden answer model inside retrieval. External web is not enabled in this connection.