Skip to main content

​
Context Settings

The Context tab controls how your assistant manages memory and shapes its output. In order, the controls are: Smart Context, AI Prompt, Context Limit (tokens), Message History Limit, Maximum Turns, and Output Schema.

​
Smart Context

Default: enabled. An intelligent memory system that includes relevant conversation history while filtering out information that isn’t useful for the current response, so the assistant isn’t overwhelmed by irrelevant history as a conversation grows.
Keep Smart Context enabled for most use cases — it’s a filtering step with no separate cost.

​
Advanced sub-panel

Turning on Advanced Mode (Settings → Advanced) reveals a nested panel with finer-grained memory controls: a summary model and threshold for when older messages get compressed, how many recent messages are always kept unsummarised, and — one level deeper — a token limit, a preservation strategy (none / fixed window / full thread), a window size, which message types count, and separate scoring weights for keyword, vector, and AI-based relevance.
Trigger After X Messages and the entire nested “Advanced” sub-block beneath it (token limit, preservation strategy, window size, message types, scoring weights) are not currently wired to anything — changing them has no effect on the assistant’s behaviour today.

​
AI Prompt

Default: off. When enabled, it takes your trigger prompt and the assistant’s behaviour text and rewrites both using a fast model from the same provider family, before the assistant runs. It does not read your tools, output schema, or do anything specific to the model you’ve selected — it’s a prompt rewrite step, not a model-aware optimiser.

​
Context Limit (tokens)

A slider from 0 to 200,000 tokens.
This control is not currently applied anywhere in the assistant pipeline. Changing it has no effect on what the assistant actually sees — every assistant currently runs on the same internal history budget regardless of this setting. Leave it at its default; there’s no tuning benefit to changing it today.

​
Message History Limit

Default: 200 messages. The maximum number of prior messages retained for conversation context. If the token budget is exceeded, older messages are dropped first. Lower it if you want tighter cost control on long-running conversations; raise it if an assistant needs to reference further back than 200 messages.

​
Maximum Turns

Default: 6. Labelled “Maximum turn count for LLM conversations.”
This limit is not currently enforced. There’s no code path that reads it, so setting it has no effect on how many turns an assistant actually takes.

​
Output Schema

A list of field name → description rows. Each description is free text — write what the field should contain and its type, e.g. "Full name of the requester (string)" or "one of: pending, approved, rejected" for an enum-style constraint. Keep descriptions short and concise, since they’re read by the model itself.
There is no validation or automatic retry. An Output Schema is turned into natural-language formatting instructions appended to the assistant’s prompt — nothing checks the response against it, and nothing retries if the assistant doesn’t follow it. Treat it as strong formatting guidance, not a guarantee.
The editor only supports flat key → description rows — you can describe a field as “a list of strings” or “an object with X and Y” in words, but you can’t build nested objects or arrays of objects as structured sub-fields. Use a schema when you need consistent field names for downstream processing — structured extraction, categorisation, or feeding another step in a flow. Skip a schema when you want natural, conversational responses — general chat, free-form writing, summarisation.

​
Best Practices

It’s a free filtering step that keeps long conversations focused. Disable only if you have a specific reason to see the full raw history.
Neither control currently has an effect. If you need to manage cost or conversation length, use Message History Limit and keep instructions and tool sets focused instead.
A vague description like "status (string)" gives the assistant little to work with. Be specific: "Current status — one of: pending, shipped, delivered, cancelled (string)".
Since nothing validates the response, test with real inputs and check the assistant is actually following the shape you described — especially for edge cases where a field might be missing.

​
Troubleshooting

Check Message History Limit isn’t set too low, and confirm Smart Context is enabled rather than disabled.
Since there’s no automatic retry, this needs a fix at the source: simplify the schema, make field descriptions more specific, mark optional fields clearly, and check the assistant’s instructions don’t conflict with what the schema asks for.

​
Next Steps

General Settings

Configure identity, timeout, and response mode

Provider Settings

Choose an AI model

Knowledge

Add documentation and reference sources

Tools & Connectors

Connect data sources and APIs