One voice note, three requests: how we built an LLM router

Описание и применение

Система автоматически разбивает голосовую заметку, содержащую несколько запросов, на отдельные задачи и направляет их соответствующим агентам, запрашивая уточнение только при необходимости.

Задача

Users record a single voice note that contains multiple requests, but the system must separate and route each request to the appropriate agent without requiring the user to issue them separately.

Результат

The system enables field workers to issue complex multi‑action voice notes in a single utterance, automatically routing each request to the correct agent and only prompting for clarification when necessary, thereby reducing interruptions and improving workflow efficiency.

Рабочий процесс

• 1. Receive the raw voice note and transcribe it.
• 2. The router parses the transcription, splitting it by distinct requests rather than sentences.
• 3. Each request is passed to the agent that declares responsibility for that type of work.
• 4. The model receives the original wording and contextual project data (timestamps, current time, card contents).
• 5. The model selects the relevant card(s) based on the search results provided by code.
• 6. If the model cannot disambiguate (e.g., multiple matching cards), the router asks the user a clarifying question; otherwise it proceeds without confirmation.
• 7. The agent executes the action and logs the change in the card’s history.

Доказательства

The router splits a message by the things asked, not by sentences.

Источник

The router splits a message by the things asked, not by sentences.

Источник

In practice, users accept 93% of permission prompts anyway.

Источник