Context windows, upload caps, and timeouts

Every chat turn sends the whole conversation again, and a configurable policy decides which earlier blocks go with it, so the turn fits. You set how many images are re-sent, and how many pieces of encrypted content go back to the original encryptor.

The limits

Limit	The number	Whose limit it is	How to get past it
Request size	About 4.5 MB for the whole request - message, history and attachments together	The edge network in front of the hosted app - Free and Pro	Turn on Direct Connection; hide old messages with Cleanup; remove large attachments; or change the policy
Request duration	Five minutes from send to last token	The same edge network - Free and Pro	Turn on Direct Connection, or use a faster model
Context window	Whatever the model states - it is on the model's own entry in your list	The model	A model with a bigger window, or a shorter history
Reply length	The model's output limit, i.e. Maximum Tokens	The model, plus your settings	Press Continue..., or raise Maximum Tokens

The first two belong to the hosted deployment, not to the software. Run your own instance and they become whatever your proxy and runtime impose. The error text follows: the request size limit of the server there, and the edge named only on the hosted app.

A refused request comes back as a card:

Request Too Large
  **Experimental:** enable Direct Connection to <vendor> to work around size limitations.
  [ Direct Connection - Now Try Again ]

That offer appears where your model's service qualifies, and switches the conversation in place.

The badge at the bottom of the composer

Badge colour	What it means
amber	under a quarter of the model's window would remain after this turn
red	the turn no longer fits, and the number becomes how many tokens you are over

The badge counts the message you are typing, and its colour says whether the turn fits. Hover it for the arithmetic:

▶ N available message tokens

 = Model max tokens: ...
     - This message: ...
          - History: ...
     - Max response: ...

Over the limit, that heading reads excess message tokens instead, prefixed ⚠️. On phones the count is an approximation rather than an exact tokenisation. Where the model has prices attached, the badge shows the turn's cost in place of the token count. The tooltip adds the cost breakdown.

Working through a long document

1. One chat per section. A fresh conversation resets the history you re-send every turn.
2. Prune in place. The right-hand panel's Cleanup hides messages you no longer need from the request, keeping them on screen.
3. Drop attachments once you have their answer. They are re-sent every turn until removed.
4. Prefer text over page images. A PDF sent as extracted text costs a fraction of the same PDF sent as page pictures (attachments).

When this doesn't work

Each message on screen maps to exactly one of the limits above.

On screen	The limit	Where to go
Request Too Large, or **Request too large** in an exported chat	Request size	This page, above
An unexpected issue occurred: **connection terminated**.	Request duration - the edge cut a live stream	Stalled or truncated answers
An unexpected issue occurred: **network error**.	Your own connection dropped	Stalled or truncated answers
The AI provider interrupted mid-stream: **upstream dropped**.	The AI service closed the socket - not a Big-AGI limit	Stalled or truncated answers
Token limit hit. with a Continue... button	Reply length	Press Continue..., or raise Maximum Tokens

Related

• Attachments and outputs
• Model settings
• Direct Connection
• Stalled or truncated answers