Glossary
Grounding (AI)
Making a language model answer from evidence it was just shown, not from whatever it absorbed in training. Grounding is the discipline; RAG is the most common way to do it; hallucination is what happens without it.
General definition
Grounding (AI) means supplying a language model with authoritative information at answer time and instructing it to base its response on that information, ideally with citations. A model’s training data is a compressed, dated and unattributed memory; grounding replaces reliance on that memory with a reference to something the system can point to: a policy document, a product record, a search result, the output of a calculator or an API call. The term is also used more narrowly for connecting a model to live data, as in grounding with web search or with an enterprise data source.
- Document grounding: retrieve relevant passages from a corpus and place them in the prompt (RAG)
- Tool grounding: let the model call a function, database or API through tool calling and reason over the returned facts
- Structured grounding: give the model tables, records or a knowledge graph query result rather than prose
- Citations: require the answer to reference the passage or record it used, so a human can verify it
- Refusal: instruct the model to say the sources do not cover the question instead of filling the gap
The failure grounding prevents is hallucination: fluent, confident text that is not true. Grounding reduces it but does not eliminate it. A model can ignore the evidence, misread it, or blend it with training memory, so grounded systems are evaluated on faithfulness (does the answer follow from the sources?) and answer relevance, not only on whether the answer sounds right. Quality of retrieval matters as much as the model: wrong or partial passages produce a wrong but well-cited answer.
Grounding also shapes what a product can promise. A grounded assistant can be scoped to a knowledge base the organisation controls and updates, which is the only workable basis for support, clinical, financial or legal assistants where an invented figure has consequences. It moves the accuracy question from “how good is the model” to “how good and current are our sources”, which is a question a team can actually act on.
In the Ethora ecosystem
Agents in the Ethora AI SDK are grounded by default in a per-agent knowledge base built from a website crawl, uploaded files and live web indexing, and the system prompt can require the agent to answer only from those sources and to say when they are silent. Tool calling adds the second kind of grounding: an agent can look up a booking, a balance or a record in the customer’s own systems before it replies rather than guessing.
Because retrieval and the model run wherever the deployment does, a dedicated or self-hosted Ethora install keeps the source documents, the index and the transcripts inside the customer’s environment. Every grounded reply is an ordinary message in the room, retained under the configured policy and visible in the audit trail, so a reviewer can trace an answer back to the sources the agent had at the time.
