LLM Context and Tokens
Token limits control how much text and tool data Assistela may send in one model context; they are not a message count or the price of one answer.
What a token means here
A token is a small unit of text processed by an LLM. The context includes the current request, recent chat, memory, temporary information, tool descriptions, and tool results. A longer context lets the model see more history, but can be slower and more expensive.
Maximum usable context is not the “number of tokens in one answer.” It is the budget for the complete working input area Assistela may use. Some space is automatically reserved for model output.
Where to change the limit
Open the AI account
Choose Assistela → Settings → AI Accounts and select the account on the left.
Open the Model tab
You will see the Detected limit reported by the model or provider and the Used limit Assistela actually uses.
Expand Manual token limits
Enter a token count in Maximum only when you know the model constraint or deliberately want to control cost.
Choose a safe value
Use safe 128K default restores the recommended ceiling. Use model maximum uses the detected technical limit. A lower value can act as a cost cap.
Save, check, and test
Save the account, run its connection check, and test a longer conversation with a tool you normally use.
The model loses detail sooner, Assistela compresses history more often, and complex work with several tools may not fit. Do not use expert-only minimum values for a normal account.
Automatic history compression
Under Settings → Advanced → App Operation, the When older chat is summarized table shows how compression works. Assistela shortens older history before the full limit is reached, retains some recent conversation, and reserves output space. The approximate threshold depends on the model context size.
The same page shows limits for Temporary Information, Personal Assistant email, Calendar days, and Reminder days. These are shares of available context. A higher number is not always better: more old email can crowd out the current request or tool results.
Tool schema and work steps
- Efficient / progressive loading: sends the tool list first and loads full details as needed. It is recommended and saves context.
- Detailed / full schema: sends every tool description immediately. Use it only when a specific model does not recognize tools in progressive mode.
- Work step limit for one message: controls how many partial steps Assistela may take. A higher limit does not make an answer smarter; it only permits a longer run.
Other model options
- Use model reasoning: enable only for a model that supports reasoning. Higher effort can improve quality while increasing time and cost.
- This model supports text only: enable when it cannot process images; Assistela will not send image data.
- Require a tool call when the model promises action: helps with models that sometimes claim they acted without calling a tool.
- Summarize older tool output: keep enabled so long old results do not fill context.
Keep the safe 128K default, progressive schema loading, and automatic compression. Change limits only for a specific model, failure, or budget, and verify that Assistela still follows the conversation correctly.