Page History

Reference for the ChatCompress Display action — summarises the current conversation transcript via one LLM call and atomically replaces it with a single synthetic message, preserving semantic continuity while reducing token cost.

AI Integration → Local AI → ChatCompress Action Reference

Purpose

ChatCompress is a Display Action that condenses a long conversation history into a single compact summary. It makes exactly one LLM call — passing the full transcript plus a system-level summarisation prompt — then atomically replaces the in-memory transcript with a single synthetic assistant-role message whose body is the LLM-produced summary. Subsequent ChatRequest turns see only that summary as their prior context.

Use it when a conversation has grown long enough to push against the model’s context window, or when you want to reduce token consumption on follow-up turns without discarding conversational context entirely.

When to use

• Long conversations — after many turns, transcript length grows. Compress before the model’s context window fills up and earlier turns start getting dropped silently.
• Shift continuity — compress at shift change to produce a concise handover summary the next operator can read, while giving the model a compact context baseline.
• Token cost control — on hosted or metered models, compressing a 50-turn transcript to a single summary message substantially reduces per-query token spend.
• After a topic switch — compress to carry a “what we discussed” baseline without dragging the full prior turn-by-turn exchange into the new topic.

How it works

When fired, ChatCompress executes the following sequence for the resolved (clientGuid, chatName) pair:

1. The target TChatSession control is resolved via the Object field (explicit name) or a visual-tree walk (auto, if Object is empty).
2. Short-circuit check: if the transcript contains fewer than 2 messages, the action returns status="ok" immediately without calling the LLM. Compressing a one-message or empty transcript is a no-op.
3. The full transcript is serialised and sent to the LLM with a system-level summarisation prompt asking for a concise synthesis of the conversation.
4. On a successful LLM response, the transcript is atomically replaced with a single synthetic message (role: assistant) whose body is the summary text.
5. On LLM failure, timeout, or empty summary, the original transcript is preserved unchanged and the action returns status="error". No partial replacement occurs.

The next ChatRequest turn sees only the single summary message as prior context. The full original turn history is not recoverable after a successful compress — use ChatClear if you need a reversible reset.

Configuration

On a Button (or any clickable control), add an Action dynamic with:

The Action editor hides fields that do not apply to ChatCompress. The Query field is hidden; Object, Return, and Result/Expression rows surface.

Reply envelope

The reply envelope follows the same JSON schema as ChatRequest and AI.Execute — see Local AI Reply Envelope Schema for the full field reference. Key fields for ChatCompress:

{
  "text":      "<LLM-produced summary on success, or '' on error>",
  "status":    "ok | error | disabled",
  "toolTrace": [],
  "latencyMs": 1840,
  "warnings":  []
}

On success (status="ok"), text carries the summary that was written into the transcript as the replacement message. On failure (status="error"), text is empty and the original transcript is intact. toolTrace is always empty — compress does not dispatch platform tools. latencyMs reflects the LLM round-trip for the summarisation call.

Short-circuit cases

• Empty transcript (0 messages) — returns status="ok", no LLM call, transcript unchanged.
• Single message (1 message) — returns status="ok", no LLM call, transcript unchanged. A single message cannot be meaningfully summarised further.

Both short-circuit cases return immediately without calling the LLM, so they are indistinguishable from a successful compress in terms of the return envelope. The text field is empty for short-circuit returns.

Gates

ChatCompress checks a single gate before executing:

• SolutionCapabilities[LocalAI].Enabled must be true — the master Local AI kill-switch. When disabled, ChatCompress returns status="disabled" without calling the LLM or touching the transcript.

ChatCompress does not inspect ModelOptions tool-surface bits (EnableChatHistory, EnableRuntimeMCP, per-category sub-bits). Transcript management is independent of the tool-surface configuration; the compress call uses the LLM solely for summarisation, not for tool dispatch.

Wall-clock budget

ChatCompress is subject to the same 60-second wall-clock timeout as ChatRequest. For very long transcripts, the summarisation POST may itself take several seconds. If the LLM does not respond within the budget, the action returns status="error" and the original transcript is preserved.

Target resolution

ChatCompress resolves the target chat session using the same two-path logic as ChatClear:

See also

• ChatClear Action Reference — discards the transcript entirely without an LLM call; use when continuity is not needed.
• ChatRequest Action Reference — the primary chat action that sends operator queries to the LLM and manages the transcript.
• ChatSession Control Reference — the Display control that renders the conversation thread.
• Local AI — the parent section, includes a step-by-step quick-start.
• Local AI Reply Envelope Schema — full reply envelope schema.

In this section...