voicebot — bot config

identity Name System prompt (persona / instructions)

pipeline

Form

Engine override

Forms 2/3 run on the native engines only (native_py / native_rs) — pipecat and livekit are cascaded-only.

STT — speech to text

Provider

Model

Mode (saaras only)

Language param

Expected languages (comma sep; metadata — only azure reads it, deepgram/sarvam use Language param)

LLM

Endpoint (wire)

Provider

Model

Region (blank = model default)

Max tokens

Temperature

TTS — text to speech

Provider

Voice / speaker id (pick or type)

TTS model (pick or type)

Language

Min buffer (chars) (Sarvam ws 30–200; 200 = never split before flush)

Max chunk length (Sarvam ws 50–500; blank = provider default 150)

Temperature (bulbul:v3 0.01–1.0; blank = provider 0.6; lower = steadier)

Pace (bulbul:v3 0.5–2.0; blank = provider 1.0)

VAD — voice activity (Silero)

Confidence

Start secs (attack)

Stop secs (hangover)

Min volume

End-of-turn wait (s) (lower = snappier, may cut off)

Barge-in bump (threshold raise while bot speaks; 0 = equal sensitivity, higher = echo-proof)

greeting

Mode

Reply language

Greeting text (used when mode = fixed)

recording-consent line

Spoken at the start of the call (the "this call may be recorded" disclosure). It is folded into the opener: before or appended to the greeting. Default ON for the India (DPDP) posture. An empty line is treated as OFF.

Enabled

Position

Consent text (in the bot's language)

timers

Idle stage 1 — nudge (s)

Idle stage 2 — give up (s)

Max duration soft cap (s)

Hard grace over soft (s)

call patch — doc 35 transfer to the owner

On by default for new bots (ruled 13 Aug 2026). With this off, the LLM has no transfer_to_owner tool and the Live tab's Patch to me fails with transfer is not enabled for this bot. The destination is never configured here — Django dials the owner's own user.phone (#CL-35-14, so no prompt can redirect a patch).

Enabled

Whisper the summary to the owner on pickup

Record the bridged human segment

Exit line the caller hears before the hand-off

dead-air filler — doc 44 pool

What the bot says when the real reply is slow. Off by default for fresh bots (doc 44 decision). The pool rotates and never repeats the last-used line within a call; an empty pool falls back to the single line. The 10 s stuck-tier backstop always speaks the single line, even with this off. Every line here is pre-synthesized into the audio cache on Save.

Enabled

Delay before filler (s)

Single line (pool fallback + stuck-tier backstop) Pool — one line per row (3–5 recommended)

held context unlock — mid-call staged context (docs 44/49)

Django stages the user's saved instructions that could not bind by number (staged.instructions in the pack). Scoreboard by default for new bots (ruled 13 Aug 2026; off = instant rollback). The tool gives the model fetch_context (cascaded form only); the watcher matches every committed caller turn engine-side and pushes on a clean match; both share one matcher: keyword tier first, LLM tier on a miss. scoreboard is the doc-49 context engine — deterministic commit-path matching with same-turn injection and gap questions, plus the promoted fetch_context as the post-settle door. The prompt carries name+topic labels only — content never rides the prompt.

Unlock mode

Watcher engine (llm = per-turn LLM bill)

Matcher LLM model (blank = this bot's own LLM) fetch_context wait lines (doc 44 §7 — this tool's intermediate messages)

Wait-line source

Static line (speak_before)

Progress ladder — lines while the lookup is still running

Signal → message (empty = silent signal)

keyword_miss_llm_start

match_found

close_match_need_detail (shapes the reply — usually left empty)

WhatsApp location share — doc 51

On by default. Runs only when a PINNED location instruction rides the call's pack; off = the bot only ever speaks the address aloud. Grace covers a "Hi" that arrives after the call ends.

Enabled

Post-call grace window

Grace window (min)

live rolling summary — the two-line card pair (plan 63)

On by default. A separate cheap LLM side-lane summarizes after committed caller turns and streams call.summary to the live card. Fire-and-forget: it can never delay a turn. Global kill: VOICEBOT_ROLLING_SUMMARY_ENABLED.

Enabled

Model (blank = this bot's own LLM)

Min gap between summaries (s)

Timeout (s)

Max tokens

tool wait lines — doc 44 per-tool intermediate messages

Per HTTP tool declared on this bot: where its wait line comes from, the static line, and the time-progress ladder. Progress signals are local-tools-only in v1 (fetch_context above) — an HTTP tool is one POST, one response, so it gets the time ladder only.

guardrails — per-bot screening rails

Each rail is one block of the screened-call prompt. Uncheck to drop it from THIS bot (conditional rails then never arm); edit the text to override it. Text identical to the shipped default is stored as "no override", so default-text bots keep tracking code updates. The backend save-time scans (sensitive-info flag, profanity block) are a separate floor these switches do not touch.

Loading rail defaults…

instructions — bot-level instruction text

Every instruction the system itself teaches or speaks on this bot's calls — engine addenda (end-call, transfer), screening teachings, assistant-mode prompts, wrap-up nudges and the abuse warn/close lines. These are not part of the persona prompt above; they are the machinery's own words. Edit to override for THIS bot; text identical to the shipped default is stored as "no override", so default-text bots keep tracking code updates. Some texts carry machine tokens (<<END_CALL>>, {line}, tool names) — a save that drops a required token is rejected.

Loading instruction defaults…

Select a call on the left to see its transcript, audio, latency and event timeline.

call stream — GET /v1/calls/{id}/events

detached floor: — seq 0

transcript — interim · segment · turn (roles: caller / assistant / owner)

Attach to a call to watch it speak.

operator door — doc 34 steer / say-exact · doc 35 patch

Kind

Door

Text (optional for patch — it becomes the transfer summary)

Delivery is confirmed on the stream as message.status: delivered → acting → done/failed.

typed-caller harness — text enters the pipeline AS the caller Caller says

Only works while the call is live in THIS process (the barge gate and endpointing engage normally).

account stream — GET /v1/events (presence across a user's calls)

User id (required for an operator token)

Bot id filter (optional)

detached

raw frames

who / which call

Acting user id (Django admin → Users → copy id)

Session id (= the voicebot call id)

station 2 — the gate (THE decision of record)

Caller number

CNAP name (optional)

Pre-answer branches on status alone: 200 = connect to the bot, 409 = ring through untouched.

escalation ladder + hold-and-ring

Tier

Urgency class (hold-and-ring overrides this)

Claim line (what the caller says they want)

Hold and ring

Ladder id (auto-filled by Start)

Ring / receipt action

Message text (required when the action is "message")

The 3-button ring door is the app plane (needs impersonation); receipts ride the same ladder id.

call patch — doc 35 transfer chain

Provider CallSid

Door

Summary (what the owner is told at ring)

Whisper on pickup

Switchman answers "should this call continue?" after the first applet exits; the Connect payload carries the ring cap, hold tone and record flag. The re-fetch carries DialCallStatus — it must never offer a second destination, and it settles the patch so the failure branch owns the call.

write-back — seed a call record so the Records tab has something to show

Caller number

Kind (axis 1, at end)

Transcript (one line each, "role: text") Summary

acting user User id

Everything here is the app plane (/api/v1/calls/context) — it needs VOICEBOT_CONSOLE_IMPERSONATION=true on the voicebot. This tab renders ONLY contract fields; tokens/questions are engine internals and never on the wire.

save an instruction — the one write door (save pass runs on every save) Instruction, in the user's words (the only required field)

Person name (optional)

Person number (optional — becomes a HARD lock: only this caller can ever unlock the row)

Topic (optional)

Expires at (optional — otherwise derived from the text / mode)

instructions on file

Set an acting user, then Reload.

blocked words — save-time profanity fuse (global)

A save whose text contains any of these words is refused outright (422), before the LLM pass — not confirmable, unlike the sensitive-info flag. The list is deployment-wide, takes effect on the very next save, and also blocks instructions that merely mention a listed word — the save-pass prompt clause handles wording the list misses.

Set an acting user, then Reload.

acting user User id

Every call on this tab is the app plane — it needs VOICEBOT_CONSOLE_IMPERSONATION=true on the voicebot.

setup health · onboarding · heartbeat

MCC+MNC (SIM carrier, e.g. 405840 for Jio)

verification probe (outbound class 4)

Mode

Probe id (auto-filled)

focus · decline telemetry · contacts

Focus until (blank clears focus)

Declined number

Contacts (one per line: "+919876543210, Name" — this REPLACES the whole phone book)

context — what the screening brain is allowed to know

Add

Expires at (required for an expectation)

Bound number / scope (optional)

Text

deactivation — the exit ceremony

calls tab shell — GET /api/v1/calls/tab

selected call card

Pick a record on the left.

Teach

Promote a caller claim to a durable fact

people & search

Person by number

Search

notifications (N1–N12)

voicebot health & console wiring

Django ops rollup — GET /internal/voicebot/ops/metrics

Window (hours)

outbound classes (1 redial · 2 demo · 3 talk-to-assistant)

Acting user id

Session id to redial

All three refuse loudly when the call_outbound kill switch is off — that refusal IS the test.

raw Django console — anything the buttons above don't cover

Method

Plane

Path after the plane prefix JSON body (ignored for GET)