citeformer¶

A bulletproof way to generate verifiably cited text from language models.

citeformer makes citation fabrication structurally impossible at the logit level when you use a grammar-level constrained-decoding backend. The model can only emit citation markers that refer to sources you’ve actually supplied. Reference lists are rendered deterministically by home-grown formatters (APA 7, MLA 9, Chicago author-date, IEEE, Nature, Vancouver) — the model never touches the bibliography.

Important

citeformer is pre-1.0 and under active development. Ten backends ship today: three local (HF transformers, vLLM, llama.cpp) that mask in-process; Fireworks (native GBNF — citeformer’s cite-id grammar drops in unchanged); and six strict-structured-output API backends (OpenAI, Anthropic, OpenRouter, Together, Gemini, Mistral) — strict structured outputs is now real token-level constrained sampling on every modern provider. See the architecture doc for the per-provider breakdown.

Install¶

# Minimal — no backends, just the core types.
pip install citeformer

# With the HF transformers backend (local, logit-layer enforcement).
pip install 'citeformer[hf]'

# With an API backend.
pip install 'citeformer[openai]'    # or: anthropic, openrouter, fireworks, together, gemini, mistral

# Everything cross-platform.
pip install 'citeformer[all]'

Python 3.11+ (tested up to 3.14).

Get started¶

Getting started — install, quickstart, verification.
Guarantees — what “bulletproof” actually means, tier by tier.
Verification — how NLI-based verify() works.
Reference — architecture, the three §10 contracts, releasing.
Historical spec — the frozen genesis plan this was built from.

Why citeformer exists¶

LLM-generated citations are wrong 14–95% of the time depending on the benchmark; RAG systems still fabricate 3–13% of cited URLs; NeurIPS 2025 accepted ~50 papers with AI-generated fake references. Prompting doesn’t fix it. Post-hoc verification doesn’t fix it. The only real fix is structural: make the invalid output token-impossible before the model even reaches that decision point.

That’s the jsonformer insight, applied to citation attribution in RAG. citeformer wraps modern constrained-decoding libraries (XGrammar, llguidance) plus six hand-written CSL formatters (APA 7, MLA 9, Chicago author-date, IEEE, Nature, Vancouver — see ADR-004 for why we rolled our own) into a single API where the guarantee is honest: given a supported local backend, a citation marker pointing at a non-existent source is a logit-level impossibility, not a hope.