Token usage — what counts

Your dashboard shows token usage as two colours — pre-processing and generation. Here is what each one is, and what does and doesn't cost you tokens.

The two kinds of tokens

A large language model bills for the text it reads and the text it writes, measured in tokens (roughly ¾ of a word each). agent4.io splits your usage the same way:

• Pre-processing (input). Everything handed to the model so it can answer:
  • the person's message,
  • the recent conversation history,
  • the relevant chunks retrieved from the agent's knowledge bases,
  • and the agent's own instructions — its soul (system prompt), task and any loaded skill.
• Generation (output). Everything the model produces:
  • the reply your customer reads,
  • and, when the agent uses a tool or skill, the structured call it writes to trigger it.

For a grounded agent the pre-processing bar is usually the larger of the two, because retrieved knowledge is added to every question.

What does not cost tokens

• Retrieving and storing documents. Building and searching a knowledge base does not consume your quota — only the text the model actually reads does.
• Delivering messages. Sending and receiving over the web widget, Telegram or WhatsApp is free of token cost.

Reading your usage

On the dashboard overview:

• Token usage over time — one bar per day (or hour), split into pre-processing and generation, with a 24h / 7d / 30d range toggle. Use it to spot a spike.
• Token usage by agent — the same split, one row per agent, widest first. Use it to see which agent is doing the most work.

All of this counts against your monthly quota; when it runs low, see Quotas & usage.

Keeping usage down

• A focused agent answers in fewer, shorter turns than a do-everything one.
• Retrieving fewer chunks per question lowers input tokens; the default is already tuned, so raise it only if answers are missing detail.