Context

Overview

Effective AI assistants require relevant context without exceeding token budgets or diluting model attention. Asiyst uses a multi-layered Context Stack that dynamically balances system instructions, user metadata, live DOM context, and retrieved knowledge.

The Context Stack

Every LLM inference turn receives a curated prompt consisting of:

1. System Persona: The global personality, constraints, and instructions configured in your assistant settings.

1. Current Page Context: The active URL path, page title, and query parameters where the visitor is located.

1. User Profile & Traits: Identified user traits (planTier, role, company).

1. Knowledge Retrieval Chunks: The top 3–5 document excerpts retrieved for the specific user query.

1. Conversation History: The recent dialogue turns between user and assistant.

Sliding Window & Turn Compression

To keep latency low and avoid token limits during long conversations:

• The last 6 message turns are preserved with full verbatim text.

• Older conversation turns (turns 7+) are automatically compressed into an executive summary by a lightweight background model.

• This ensures the assistant remembers agreements made 20 messages ago without wasting context window tokens.

Dynamic Page Context Injection

You can inject custom page-level variables into the assistant's active memory:

import { Asiyst } from "@asiyst/sdk";
// Inject active product context on an e-commerce page
Asiyst.setContext({
  productId: "prod_88291",
  productName: "Enterprise Cloud Server",
  inventoryStatus: "In Stock",
  currentPrice: "$499/mo",
});

Token Budget Allocation