Knowledge bases

A knowledge base is a dataset for AI retrieval. Put processed tables and files into scope and add usage instructions. When you ask a question, AI retrieves and answers within that scope, and the answer can be traced back to the source. This page covers the path from creating a knowledge base to asking questions.

Your data

AI participation

Processed tables and files are put into a retrieval scope with usage instructions.

For each question, AI retrieves within scope and answers from the retrieval result and instructions, with citations.

Create a knowledge base

A knowledge base has two parts:

Part

Purpose

Selected data

Defines the AI retrieval scope: the tables and files belonging to this knowledge base.

Usage instructions

Tell AI what this data is, its conventions, and the questions it is suitable for.

Both Q&A and agents retrieve from knowledge bases. AI does not retrieve data outside the scope.

Usage instructions directly affect answer quality

Usage instructions are not notes: AI reads them while retrieving and answering. State three things clearly:

  • What the data is: source, scope, and time span, for example, “Sales contracts from 2024–2025, including scanned copies.”

  • Conventions: information such as field meanings, units, and abbreviations that can be misread.

  • What it is and is not suitable for: questions this data can and cannot answer.

Vague instructions degrade both retrieval relevance and an answer’s boundaries. Stating these three things clearly is the lowest-cost way to improve quality.

Ask questions

  1. Your question becomes a retrieval request that finds relevant content within the knowledge base scope.

  2. Retrieved snippets and usage instructions are sent to the model together.

  3. The model answers from those materials. Citations in the answer can be traced to the source through data lineage.

Validate a new knowledge base with questions whose answers are in the source, to check answer and citation accuracy, and questions whose answers are not in the source, to check that AI truthfully says nothing was retrieved instead of inventing an answer.

Trace back by segment when an answer is unsatisfactory

Symptom

Check first

Go to

The answer says there is no relevant information, but the source clearly has it.

Was the file processed? Were the outputs added to the knowledge base?

Data processing

The answer cites irrelevant content.

Is the knowledge base scope too broad? Do the instructions state suitable questions?

Earlier on this page

The answer uses the wrong convention.

Do the instructions state the conventions?

Earlier on this page

You need to confirm which file a statement came from.

Trace lineage from the answer citation.

Data Catalog

Next steps

What you want to do

Go to

Have an agent use this knowledge base to run tasks

Agents

Walk through the complete path from upload to Q&A

Quick start

For AI agents: read a knowledge base’s usage instructions before retrieval. Do not fill gaps with data outside scope. Answers must use retrieval results and cite them. If retrieval is empty, say so truthfully.

Last updated on