Runtime Configuration

Language hints (comma)

Biases detection, e.g. en, hi.

EOT threshold

Confidence to end a turn. Lower = ends sooner (more interruptions).

EOT timeout (ms)

Silence after speech that force-ends a turn. Too low cuts mid-sentence on pauses.

Defer connect on start true false

Connect the Flux socket in the background so the greeting isn't blocked ~1s. Keep true.

Turn stop debounce timeout (ms)

Controls Pipecat's post-STT delay before triggering the LLM (NOT Deepgram eot_timeout_ms). Lower values reduce latency but may increase fragmented turns. Range 100–1000; default 500.

Enable EagerEOT telemetry true false

Stage-1 measurement only: asks Flux for EagerEndOfTurn events and records the would-be LLM head start. No speculative draft, no early speech. Default false (off = identical to today).

EagerEOT threshold

Confidence for the early EagerEndOfTurn signal. Must be ≤ EOT threshold; range 0.3–0.9. Only sent to Flux when EagerEOT is enabled. Leave blank for off.

Max tokens

Caps reply length. Smaller = faster, shorter phone replies.

Request timeout (s)

Hard timeout per LLM request (wired to the Cerebras client) and the bound on outbound greeting generation.

In-call stall timeout (s)

ARC-06: per-turn TTFB bound for the in-call LLM. A stalled request retries once on Cerebras, then fails over to the Gemini fallback model for that turn.

Top-p

Nucleus sampling. Unset = provider default; blank keeps the saved value (clear via scripts/set_llm_penalties.py).

Frequency penalty

Intra-turn repetition guard. Keep ≤ 0.5 — higher distorts Kannada/Telugu suffixes. Unset = off; blank keeps the saved value.

Presence penalty

Discourages tokens already used in this reply. Keep ≤ 0.5. Unset = off; blank keeps the saved value.

Enable speculative LLM (Stage-2) true false

Starts an out-of-band Cerebras draft on EagerEndOfTurn and releases it at EndOfTurn when the eager transcript matches the final — hiding the turn-stop debounce. Requires EagerEOT telemetry on. Default false (off = identical to today).

Language mode per_turn fixed

Recall min overlap

Content-word overlap needed to surface an earlier turn. Higher = fewer/cleaner recalls (1–6). An explicit “you told me…” relaxes it to 1.

Memory block token cap

Budget for facts + recall + summary injected each turn, so a big block can’t inflate latency (100–2000).