How to reduce Claude hallucinations by grounding answers in your documents

A hallucination is a confident answer that is not supported by anything real. The most reliable fix is grounding: give the model the source material in the request and tell it to answer only from that material. The Anthropic API has a built-in citations feature that ties each sentence of the answer back to a span of your document, so you can verify claims instead of trusting them.

• The Anthropic SDK installed and an API key set
• One or more source documents (text or PDF)
• A question you want answered strictly from those documents

Step 1: Put the source in a document block

Instead of pasting the source as plain text, wrap it in a document content block. This lets the API track character positions for citations later.

from anthropic import Anthropic

client = Anthropic()
source = open("policy.txt").read()

resp = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=1024,
    messages=[{
        "role": "user",
        "content": [
            {
                "type": "document",
                "source": {"type": "text", "media_type": "text/plain", "data": source},
                "title": "Refund policy",
                "citations": {"enabled": True},
            },
            {"type": "text", "text": "What is the refund window? Answer only from the document."},
        ],
    }],
)

Step 2: Read the citations off the response

With citations enabled, the answer is split into text blocks. Cited blocks carry a citations array, each entry pointing at the exact span of the source the claim came from. Walk the blocks to show or log where each statement is grounded.

for block in resp.content:
    if block.type != "text":
        continue
    print(block.text)
    for c in (block.citations or []):
        print("  source:", repr(c.cited_text))

Step 3: Keep the model honest about gaps

Test with a question the document cannot answer. A well grounded setup should refuse to invent. If it still makes something up, tighten the instruction and confirm the source actually contains the section you expected.

Result: with the source in a document block and citations on, the refund answer pointed straight at the 30 day clause, and the gift card question returned an honest I-do-not-know instead of a fabricated policy. You can now audit every claim.