Grounding
A technique that secures the accuracy of AI model outputs by linking them to reliable external data or specific knowledge bases. By allowing the model to reference up-to-date information or internal data not included in its training, it suppresses hallucinations and provides grounds for the responses.
Detailed explanation
Why it matters for tool selection
When choosing enterprise AI tools, grounding performance is the benchmark for determining 'business feasibility.' Just as important as the model's intelligence itself is how accurately it extracts and reflects in-house data or real-time web information in its answers. Poor grounding can lead the AI to tell plausible lies, causing critical errors in business decision-making.
What to check when choosing tools
- Does it provide clear source citations with links for the answers?
- Does it include real-time web search capabilities or integration with internal databases (ERP, CRM, etc.)?
- Is the security and privacy of the data used for grounding guaranteed?
- Does it provide a 'Grounding Score' that measures how relevant the retrieved information is to the answer?
Example
A typical example of grounding is when a user asks, 'What is our company's operating profit for the first quarter of this year?' and instead of reasoning based on general knowledge, the AI searches the company's latest financial report PDF, answers 'The Q1 operating profit is 5 billion won,' and cites page 5 of that document as the source.
Commonly confused terms
Fine-tuning
The process of retraining the model's 'brain' to fit a specific domain, essentially having it 'memorize' new information.
Grounding
The process of giving the model an external 'reference book' and having it answer based on it, which is more advantageous for reflecting real-time data.