Grounding

Definition

Grounding is the practice of connecting an LLM’s responses to verifiable, external sources of truth — so the model answers based on real data rather than what it “memorized” during training.

When an LLM answers a question from memory alone, it might hallucinate — generating plausible-sounding but wrong information. Grounding prevents this by anchoring responses to specific documents, databases, search results, or APIs. If the model says “Claude costs $3 per million input tokens,” grounding means there’s an actual source document backing that claim.

RAG (Retrieval-Augmented Generation) is the most common grounding technique, but grounding is the broader concept. It includes any method that ties model output to verifiable information: web search, database queries, API calls, or document retrieval.

How It Works

UNGROUNDED (risky):
  User: "What's the latest Claude pricing?"
  Model: [relies on training data from months ago]
  Model: "Claude Sonnet costs $3 per million input tokens" ← might be outdated

GROUNDED (reliable):
  User: "What's the latest Claude pricing?"
  System: [retrieves current pricing page]
  Model: [reads the retrieved document]
  Model: "According to Anthropic's pricing page, Claude Sonnet 4.6
          costs $3 per million input tokens." ← sourced from real data

Grounding Methods

1. RETRIEVAL GROUNDING (RAG)
   Query → Search vector database → Relevant docs → Model reads & answers
   Example: Company knowledge base Q&A

2. SEARCH GROUNDING
   Query → Web search API → Search results → Model synthesizes
   Example: Perplexity AI, Google Gemini with Search

3. TOOL GROUNDING
   Query → Model calls API/tool → Real-time data → Model responds
   Example: "What's the weather?" → calls weather API → real answer

4. DOCUMENT GROUNDING
   Upload document → Model reads it → Answers only from document
   Example: "Summarize this contract" with uploaded PDF

Why It Matters

• Accuracy — Grounded responses are verifiable against their sources, dramatically reducing hallucination
• Freshness — Models can access information published after their training cutoff
• Trust — Citations let users verify claims themselves
• Enterprise adoption — Most companies won’t deploy LLMs without grounding because the risk of wrong answers is too high
• Legal/compliance — In regulated industries, every AI-generated claim may need a traceable source

Example

# Grounding with web search using Perplexity-style approach
from anthropic import Anthropic

client = Anthropic()

def search_web(query: str) -> list[dict]:
    """Simulate web search — in production, use a real search API."""
    # Use SerpAPI, Brave Search, or Tavily in production
    return [
        {
            "title": "Claude Pricing - Anthropic",
            "url": "https://anthropic.com/pricing",
            "snippet": "Claude Sonnet 4.6: $3/M input, $15/M output. Claude Opus 4.6: $15/M input, $75/M output."
        },
        {
            "title": "Claude API Documentation",
            "url": "https://docs.anthropic.com",
            "snippet": "Claude supports 200K token context windows across all models."
        }
    ]

# Step 1: Search for relevant information
query = "What are the current Claude API pricing tiers?"
search_results = search_web(query)

# Step 2: Ground the model's response in search results
context = "\n\n".join(
    f"Source: {r['title']} ({r['url']})\n{r['snippet']}"
    for r in search_results
)

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=500,
    temperature=0,
    system="""Answer questions using ONLY the provided search results.
Cite sources with [Source Title](URL) for every factual claim.
If the search results don't contain the answer, say so.""",
    messages=[{
        "role": "user",
        "content": f"Search results:\n{context}\n\nQuestion: {query}"
    }]
)

print(response.content[0].text)
# → "According to [Claude Pricing - Anthropic](https://anthropic.com/pricing),
#    Claude Sonnet 4.6 costs $3/M input tokens and $15/M output tokens..."

# Document grounding — answer only from uploaded content
def grounded_qa(document: str, question: str) -> str:
    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=500,
        temperature=0,
        messages=[{
            "role": "user",
            "content": f"""You are a document analysis assistant.
Answer the question based ONLY on the document below.
If the answer is not in the document, respond: "This information is not in the provided document."

--- DOCUMENT START ---
{document}
--- DOCUMENT END ---

Question: {question}"""
        }]
    )
    return response.content[0].text

# The model can only use information from the document — no hallucination
contract = open("service_agreement.pdf.txt").read()
answer = grounded_qa(contract, "What is the termination notice period?")

Grounding vs RAG

RAG is the most popular grounding technique, but grounding also includes web search, API calls, and tool use.

Key Takeaways

• Grounding connects LLM responses to verifiable external sources, reducing hallucination
• RAG is the most common grounding technique, but web search and API calls also qualify
• Grounded responses include citations, letting users verify claims
• Essential for enterprise AI deployment — most companies require grounding for production LLM applications
• Perplexity AI and Google Gemini with Search are prominent examples of grounded AI products

Part of the DeepRaft Glossary — AI and ML terms explained for developers.