Limits and plans

What we count

  • Runs — every start of a flow: webhook, schedule or manual. A sub-flow is not counted separately: the whole chain is one run.
  • AI operations — calls from LlmPrompt, Agent and the flow assistant. One call can cost several operations when the prompt is large.
  • Stored vectors — what your knowledge bases turn into once they are indexed.

The cap on AI input

Prompt length is capped by plan. A prompt that does not fit fails the node with a clear message instead of silently truncating the text.

With your own key there is no cap. Connect your provider in the profile: then you pay your provider for tokens and the AI quota is not spent.

When a quota runs out

The plan's quota comes back on the first of the month. If it ran out earlier, changing plan is not the only way — billing settings let you top up beyond the plan:

  • flow runs — 5,000, 25,000 or 100,000. This is what a busy site spends fastest: every /_api/* call is a run.
  • AI operations — 500, 2,500 or 10,000.

What you buy does not expire at the end of the month and is spent only after the plan is used up. The two kinds stay apart: runs never pay for AI, and AI never pays for runs.

For AI there is one more way, and it is free: connect your own provider key — then the operations quota is not spent at all.

Vector storage

The first two counters reset on the first of the month. Vectors do not: a knowledge base keeps them until you delete it, so the plan caps how many you hold at once, not how many you make per month.

How many vectors a document takes depends on the chunk size: by default one vector per ~800 characters of text. The current figure sits on the dashboard next to runs and AI operations.

When there is no room left, new indexing stops with a message while existing search keeps working. There are three ways out:

  • delete a knowledge base you no longer use;
  • move to a larger plan;
  • connect your own Qdrant in the settings — datasets on your own store are not counted against the plan at all.

Platform limits

  • one run — at most 1000 steps
  • sub-flow nesting — three levels
  • history — the last 10,000 runs per flow
  • database select — up to 1000 rows per request
  • webhook wait — up to 30 seconds
  • knowledge base chunk size — between 200 and 3000 characters

Files

The size cap depends on the kind: images, video, audio, PDF, text and spreadsheets differ. A file can be made public, which gives it a direct link; video and audio are served in ranges so a player can seek.