Agent

Model Intelligence

How Katexs routes reasoning, memory, and latency budgets during a live conversation.

Turn pipeline

┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│  Customer   │────▶│   Phone or  │────▶│   Katexs    │
│  Calls or   │     │    Chat     │     │    Agent    │
│  Messages   │     │   Widget    │     │  (AI Brain) │
└─────────────┘     └─────────────┘     └──────┬──────┘
                                               │
                    ┌──────────────────────────┼──────────────────────────┐
                    ▼                          ▼                          ▼
              ┌─────────┐              ┌─────────┐                 ┌─────────┐
              │ Calendar│              │   CRM   │                 │ Transfer│
              │  Book   │              │  Push   │                 │  Human  │
              └─────────┘              └─────────┘                 └─────────┘

Transcribe — Streaming speech-to-text with endpointing tuned for phone audio.
Reason — The LLM reads your prompt, recent turns, and retrieved knowledge chunks.
Act — If a tool is needed the agent calls it, waits for the result, then speaks.
Speak — Text-to-speech streams the first sentence while the rest is still generating.

Latency budget

Stage	Target	Notes
Transcription	< 300 ms	Streaming partials
Reasoning	< 700 ms	Model dependent
Tool call	< 900 ms	Calendar and CRM round trips
Speech	< 400 ms	First audio chunk

Memory

Each conversation keeps a rolling window of turns plus a structured summary of collected fields (name, phone, intent), so long calls never lose the details you need.