VoiceAgent Control Layer

Dashboard

VoiceAgent Control Layer

Live

Today: 0 calls Avg: 0ms Cost: $0.00 Errors: 0%

Listening for middleware traffic

Newest turn at the top · reloads from DB after deploy

Identity

Loading…

Classify

Loading…

Sticky inject

Loading…

Prompt cache

Loading…

Signals

Loading…

Clarify recovery

Loading…

Live trend

Cache fill plus stacked wall-clock (classify + KB + assemble + LLM stream). Bars are time-to-last-token, not TTFB. Pink line is first spoken chunk to 11L (TTFT).

Cache fill and stacked turn latency. Pink line = first out (TTFT).

Prompt cache

—

Token cache Hit rate

Turn latency (stacked)

—

Groq classify KB / orch Assemble LLM stream line = first out to 11L (TTFT)

Cache by model

Token cache from the served model · Requesty gateway HIT/MISS when present

Source	Hit	Tokens	Samples
Loading…

Recent turn decisions

—

When	Agent	Call / Session	Stage	Outcome	Reason	ms
Loading…

Learned intents

—

Held classifier hits only · exact phrase recall skips Groq · 30-day decay · delete is hygiene, not a correction

Phrase	Group	Hits	Seen	Agent
Loading…

Settings

Changes apply to the next request

LLM Provider Select which AI provider to use. Each has different models, pricing, and capabilities. Perplexity supports web search.

Default model Used for dashboard chat and for agents without their own mapping. ElevenLabs Model ID still overrides when it is a real model name (e.g. sonar-pro).

Max Tokens Maximum response length. Lower = shorter answers, faster, cheaper. Default 1024.

System Prompt Mode How your prompt interacts with the ElevenLabs agent prompt. Off = pass-through, Override = replace it, Prepend/Append = combine both.

System Prompt Custom prompt text merged into the request per the mode above.

Rewrite ElevenLabs policy clauses Ersetzt das generische Verbot von Rechts-/Finanzberatung und urheberrechtlich geschützten Namen. Procedures, TTS-Regeln und Character bleiben. Standard: aus.

Reasoning Effort How much the model "thinks" before answering. Higher = better quality but slower and more expensive. Only affects reasoning models.

Classifier model Themen-Classifier (Groq). Default openai/gpt-oss-20b. Per-agent override is possible via API.

Groq region Applies immediately on Save — classifier and Groq LLM, no Railway redeploy. EU needs Groq to grant regional API access; without it you get 403. —

Classifier timeout (ms) Hard skip. On timeout the call continues without a topic. Default 800 (inside the 11Labs 3 s first-token window).

Per-agent models

Same middleware URL. Each ElevenLabs agent_id can use a different LLM.

Go-live reset

Wipe test traffic before the first live agent. Type RESET in the confirm dialog. Does not clear the LLM provider prompt cache.

Reset statistics

Deletes stored telemetry so Decisions and header stats start at zero. Use this when the first live agent goes on the line.

Loading counts…

Flush KB RAM cache (manifest/docs). Next turn refetches from agent_dash. Forget learned intents (phrase overlay from test calls). Clear sticky topics (running calls will re-classify).

Type RESET to confirm

VoiceAgent Control Layer

Live trend

Prompt cache

Turn latency (stacked)

Cache by model

Recent turn decisions

Learned intents

Reset statistics

Session Detail