WHITEBOXX / GLOSSARY

Grounding: tying AI answers
to sources you trust.

Grounding means connecting an AI model’s answer to specific sources, such as your documents, databases or approved data, so that its claims come from that material and not only from what the model learned in training. A grounded answer can be checked against its sources.

By Gautam Mehra · Published · Updated

What is grounding?

Language models produce fluent text from patterns learned in training. They have no built-in link to your data or to events after their training date. Grounding supplies that link: the system finds the relevant material and asks the model to answer from it.

The most common method is retrieval-augmented generation (RAG), described by researchers at Facebook AI Research in 2020. Agents add a second method: calling tools, such as a database query or a calculation, and answering from the result.

HOW IT WORKS

How grounding works.

  1. 01

    Find the material

    Search documents or query systems for the information the question needs.

  2. 02

    Give it to the model

    Pass the relevant passages or results to the model with the question.

  3. 03

    Answer from the material

    Instruct the model to use only the supplied material and to cite it.

  4. 04

    Check the answer

    Test that each claim is supported by the source it cites.

Does grounding stop hallucinations?

Grounding reduces errors but does not remove them. A model can misread a source, combine two sources wrongly or cite a passage that does not support the claim. A citation shows where the model looked, not that it read the source correctly.

Grounding also depends on the sources. If the material it finds is out of date or incomplete, the answer will be too.

How Evidence OS grounds results.

Evidence OS agents answer from data they retrieve during the run, from 562 sources and your connected systems. Their results carry the sources and the calculations behind each finding.

How to check AI research