Your RAG Gives Answers. These Lines Make It Cite Its Work.

Tag retrieved chunks with IDs, ask Claude to cite them inline, and return a sources list — your RAG answers become verifiable, not just plausible.

What you will be able to do after this:

• Return [1] / [2] inline citations in any RAG response alongside a sources list
• Verify that cited documents actually support each claim using a structured quote field
• Build the citation layer in under 50 lines on top of your existing retriever

A RAG system that returns "Refunds take 5-7 days" is useful. One that returns "Refunds take 5-7 days [1]" with [1]: Refund Policy, section 3.2 is trustworthy. That is the difference between a demo and a production tool.

Resource: RAG with Verified Citations — Trelis Research (YouTube, 34 min)

The pattern

Your retriever returns chunks. Tag each one with a sequential ID and include them in the user turn:

import anthropic

client = anthropic.Anthropic()

def rag_with_citations(question: str, chunks: list[dict]) -> dict:
    # chunks = [{"id": 1, "text": "...", "source": "Refund Policy 3.2"}, ...]
    context = "

".join(f"[{c['id']}] {c['text']}" for c in chunks)

    response = client.messages.create(
        model="claude-sonnet-5-20260801",
        max_tokens=1024,
        system="""You are a helpful assistant. Answer using ONLY the provided context.
Cite every factual claim with a bracketed ID like [1] inline.
If the answer is not in the context, say so.""",
        messages=[{
            "role": "user",
            "content": f"Context:
{context}

Question: {question}"
        }]
    )
    return {
        "answer": response.content[0].text,
        "sources": [{"id": c["id"], "source": c["source"]} for c in chunks]
    }

Structured output version (more reliable for parsing)

import json

response = client.messages.create(
    model="claude-sonnet-5-20260801",
    max_tokens=1024,
    tools=[{
        "name": "answer_with_citations",
        "description": "Return an answer with inline citations",
        "input_schema": {
            "type": "object",
            "properties": {
                "answer": {
                    "type": "string",
                    "description": "Answer text with [N] inline citations"
                },
                "citations": {
                    "type": "array",
                    "items": {
                        "type": "object",
                        "properties": {
                            "id": {"type": "integer"},
                            "quote": {
                                "type": "string",
                                "description": "Exact text from document [N] that supports the claim"
                            },
                            "source": {"type": "string"}
                        },
                        "required": ["id", "quote", "source"]
                    }
                }
            },
            "required": ["answer", "citations"]
        }
    }],
    tool_choice={"type": "tool", "name": "answer_with_citations"},
    system="Answer using ONLY the provided context. Cite every factual claim inline.",
    messages=[{"role": "user", "content": f"Context:
{context}

Question: {question}"}]
)

result = json.loads(response.content[0].input)
# result["answer"]    = "Refunds take 5-7 days [1]."
# result["citations"] = [{"id": 1, "quote": "Refunds are processed in 5-7 business days", "source": "Refund Policy 3.2"}]

The quote field in each citation forces Claude to locate the exact supporting text — surfacing any hallucination where no supporting passage exists. If Claude cannot find a quote, it cannot fabricate a citation.

Sources: RAG with Verified Citations — Trelis Research (YouTube) · Citation-aware RAG — Tensorlake · Attributing sources in RAG — apxml.com