Grounding — llm-kit

grounding

// Tying AI answers to real, verifiable sources instead of just vibes

💡 The core idea: An "ungrounded" AI answer is based on patterns from training data — you can't verify where it came from or if it's still accurate. A "grounded" answer is tied to specific sources: "According to [document X, page 3]...". Grounding makes AI outputs verifiable, trustworthy, and useful in high-stakes situations.

Compare ungrounded vs grounded responses

Question: "What is the current minimum wage in California?"

❌ Ungrounded response

The minimum wage in California is $15.50 per hour. This is among the highest in the country. It was last updated in January 2022 and applies to all employers regardless of size.

⚠️ The AI pulled this from training data — it may be outdated. No source provided. Could be wrong.

✅ Grounded response (with RAG + search)

According to the California Department of Industrial Relations [1] ca.gov, the minimum wage in California as of January 1, 2024 is $16.00 per hour for all employers. [2] ABC News reports that fast food workers under AB 1228 receive a higher minimum of $20/hour starting April 2024.

[1] California DIR — dir.ca.gov/iwc/minimumwagehistory.htm

Retrieved: Today. Excerpt: "Effective January 1, 2024, the minimum wage in California is $16.00 per hour for all employers, regardless of size."

[2] ABC News — abc7news.com/story/fast-food-minimum-wage...

Retrieved: Today. Excerpt: "Fast food workers will earn $20 minimum wage starting April 1, 2024 under AB 1228 signed by Governor Newsom."

✓ Sources are cited. You can verify each claim. Current as of today (web search used).

Click the [1] citation badges to see source details. This is what grounding looks like in practice.

Grounding vs RAG

Grounding is the broad goal: answers tied to verifiable sources. RAG (Retrieval-Augmented Generation) is one specific technique to achieve grounding — retrieve relevant documents, inject them into context, ask the AI to answer based on them. Web search is another grounding technique. Grounding = the goal. RAG/search = the methods.

GROUNDING = VERIFIABLE ANSWERSRAG + SEARCH = COMMON APPROACHES