v0.1.13 — native Anthropic Citations API (opt-in)

[0.1.13] - 2026-05-08

Added

• Native Anthropic Citations API (opt-in). New
  use_native_citations kwarg on RagPipeline.run_and_generate —
  when True (and the provider supports it), retrieved hits are
  sent as custom_content document blocks and the model emits
  claim-level citations attached to its response text. Falls
  back to the legacy [P{n}]-marker path on providers without
  support (Gemini) and on empty-hit retrievals.
  • New types: attune_rag.ClaimCitation (response-span +
    document-index + cited-text); attune_rag.providers.base. CitationDocument, CitedResponse.
  • New helper: attune_rag.format_claim_citations_markdown
    renders response text with footnote-style claim
    attribution.
  • New protocol surface: LLMProvider.supports_native_citations
    flag and LLMProvider.generate_with_citations.
    ClaudeProvider implements both. GeminiProvider declares
    the flag = False; pipeline detects this and falls back.
  • RagResult gains claim_citations: tuple[ClaimCitation, ...]
    and used_native_citations: bool (both default empty/False
    so all existing callers are unaffected).
• attune-rag query --native-citations CLI flag — opt into
  the native path from the command line. Renders the response
  with footnote citations when active.
• attune-rag-benchmark --native-citations flag (with
  --with-faithfulness) — runs a side-by-side faithfulness
  pass on both paths and prints a comparison table including
  hallucination rate, citation emit rate, and p95 latency.
• Docs: docs/rag/native-citations.md design note covering
  when to use which path, fallback behavior, open verification
  gates (V2 cache_control, V3 doc-count ceiling), and rollback.

Removed

• OpenAI provider (attune_rag.providers.openai.OpenAIProvider) and
  the [openai] install extra. Use [claude] or [gemini] instead.
  No external consumer of attune-rag was importing OpenAIProvider;
  the provider is fully removable without callsite migrations.