Voice Agent
Speak into the mic — AssemblyAI transcribes in real-time, then your question is matched against the knowledge base. Matched answers stream pre-recorded audio from S3; unmatched queries fall back to Gemini Live.
Topic
Sub-topic
Socketdisconnected
Services—
Micidle
Audio—
Session—
Prompt Review
Child context (leave everything empty for the baseline prompt)
📖 Editing guide
Cautions
- Edits are per-session only: a page reload discards them; the server prompt never changes.
- Any edit or a new "Build prompt" clears approval — re-approve after every change.
- An approved prompt replaces everything: the topic choice and the child-context form are ignored at Connect.
- Maximum 8000 characters — the server truncates anything beyond.
- Keep section tags paired and intact — the model relies on this structure.
- Never remove or weaken
<safety>— it holds the child-protection rules.
Sections
<persona>— who Rumble is (name, character, tone); safe to tweak traits/tone.<behavioral_guidance>— conversation behavior rules; adjust carefully.<output_format>— speech output constraints; edit only when testing output style.<safety>— child safety rules; do not edit.<child_context>+ language line — generated from the form; prefer changing the form and rebuilding over hand-editing.
System prompt
Edits apply to the next Connect only. Keep the <safety> section intact.
✓ Approved — used at next Connect (topic &
child context ignored)
Connect and start speaking…
Connect and speak to see timing…
Total (speech → done)
—
Event Log