Knowledge-Aware Responses

Overview

A Knowledge-Aware Response is an AI completion that derives its facts strictly from verified documentation sources rather than foundation model training data. This eliminates hallucinated pricing, outdated policies, or false product claims.

How Grounding Operates

When knowledge-aware mode is active:

1. The user's query is matched against vector embeddings of your connected website pages, uploaded PDFs, Markdown files, and GitHub repos.

1. The model receives a system constraint: "Answer only using the verified context provided below. If the answer is not in the context, state that you do not have sufficient information."

1. The model extracts specific sentences and tags each factual claim with citation indices ([1], [2]).

Citation Linking & Transparency

In the web interface, citations appear as interactive footnotes beneath assistant messages:

• Visitors can click any citation badge to view the original source document title and verified text snippet.

• If the source is a public webpage, clicking the citation navigates directly to that section of your documentation.

{
  "message": "Asiyst requires Node.js 18 or higher for SDK installation [1].",
  "citations": [
    {
      "index": 1,
      "sourceTitle": "Installation Prerequisites",
      "url": "https://asiyst.com/docs/installation",
      "snippet": "Prerequisites: Node.js version 18.0.0 or higher..."
    }
  ]
}

Handling Unindexed Questions

When a visitor asks a question unrelated to your business or not present in your knowledge base:

• The assistant does not guess.

• It delivers a customizable polite fallback response: "I don't have verified documentation on that topic. Would you like to submit a ticket to our support team?"

• An unindexed query event is logged in your Analytics dashboard under Unanswered Questions so your team can write new documentation to address the gap.

Tuning Retrieval Sensitivity

In Project Settings > Knowledge:

• Strict Mode (Recommended): Requires high semantic similarity (>= 0.80) to answer. Ideal for compliance, medical, and financial use cases.

• Flexible Mode: Accepts moderate similarity (>= 0.65) and synthesizes broader guidance.