Claude Citations — empty citations, wrong offsets, or missing grounding

import anthropic

client = anthropic.Anthropic(
    default_headers={"anthropic-beta": "citations-2025-01-14"}  # verify current beta
)

def ask_with_citations(question: str, sources: list[dict]) -> anthropic.types.Message:
    """
    sources: [{"title": "Q3 2025 earnings", "text": "..."}]
    """
    content = []
    for src in sources:
        content.append({
            "type": "document",
            "source": {
                "type": "text",
                "media_type": "text/plain",
                "data": src["text"],
            },
            "title": src["title"],
            "citations": {"enabled": True},   # opt-in per document
        })
    content.append({"type": "text", "text": question})

    return client.messages.create(
        model="claude-opus-4-7",
        max_tokens=1000,
        system="Answer strictly from the provided sources. Cite every claim.",
        messages=[{"role": "user", "content": content}],
    )

response = ask_with_citations(
    "What was Q3 revenue and headcount?",
    [
        {"title": "Q3 2025 earnings", "text": "Revenue in Q3 2025 was $4.2M ..."},
        {"title": "Ops report Q3", "text": "Total headcount at end of Q3 was 47 ..."},
    ],
)

for block in response.content:
    if block.type == "text":
        print(block.text)
        for cite in block.citations or []:
            print(f"  [Source #{cite.document_index}: '{cite.cited_text}']")

import anthropic
from typing import Iterable

client = anthropic.Anthropic(
    default_headers={"anthropic-beta": "citations-2025-01-14"}
)

def chunked_document(chunks: Iterable[str], title: str) -> dict:
    """Wrap a list of chunks as a content-typed document with citations."""
    return {
        "type": "document",
        "source": {
            "type": "content",
            "content": [
                {"type": "text", "text": chunk}
                for chunk in chunks
            ],
        },
        "title": title,
        "citations": {"enabled": True},
    }

# Build a chunked document from paragraphs
paragraphs = [
    "The company was founded in 2018.",
    "Q3 2025 revenue was $4.2M, up 18% YoY.",
    "Headcount at end of Q3 was 47.",
    "The board consists of five directors...",
]

doc = chunked_document(paragraphs, title="Q3 2025 earnings")

response = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=500,
    system="Answer from the provided document. Every fact needs a citation.",
    messages=[{"role": "user", "content": [
        doc,
        {"type": "text", "text": "What was Q3 revenue?"},
    ]}],
)

for block in response.content:
    if block.type != "text":
        continue
    print(block.text)
    for cite in block.citations or []:
        # For content-typed sources, cite includes block start/end indices
        print(f"  [doc #{cite.document_index}, block "
              f"{cite.start_block_index}-{cite.end_block_index}: '{cite.cited_text}']")

import anthropic

client = anthropic.Anthropic(
    default_headers={"anthropic-beta": "citations-2025-01-14"}
)

STRICT_SYSTEM = """
You answer strictly from the provided documents.

Rules:
1. Every factual claim must have a citation.
2. If the documents do not contain the answer, say so exactly:
   "The provided sources do not contain this information."
3. Do NOT use your parametric knowledge — only the documents.
"""

def call_with_strict_grounding(question: str, docs: list[dict],
                              max_attempts: int = 2) -> anthropic.types.Message:
    for attempt in range(1, max_attempts + 1):
        content = [
            {
                "type": "document",
                "source": {"type": "text", "media_type": "text/plain", "data": d["text"]},
                "title": d["title"],
                "citations": {"enabled": True},
            }
            for d in docs
        ] + [{"type": "text", "text": question}]

        response = client.messages.create(
            model="claude-opus-4-7",
            max_tokens=1000,
            system=STRICT_SYSTEM,
            messages=[{"role": "user", "content": content}],
        )

        # Verify: every text block with non-trivial content has ≥1 citation
        # or explicitly says the docs do not contain the answer
        acceptable = True
        for block in response.content:
            if block.type != "text":
                continue
            text = block.text.strip()
            if not text:
                continue
            has_citations = bool(block.citations)
            disclaims = "provided sources do not contain" in text.lower()
            if not (has_citations or disclaims):
                acceptable = False
                break

        if acceptable:
            return response
        print(f"[attempt {attempt}] response contains ungrounded claim, retrying")

    return response  # last attempt, even if imperfect