citeformer.backends¶
Backend implementations for citeformer.
Each backend adapts a model runtime to the common Backend ABC (see base.py).
Only the Backend ABC and MockBackend are re-exported here because they
have no optional-extra dependencies. Import the real backends directly from
their submodules — they each require the matching extra.
Local backends (logit-layer enforcement, in-process):
citeformer.backends.hf.HFBackendneedspip install citeformer[hf].citeformer.backends.llamacpp.LlamaCppBackendneeds[llamacpp].citeformer.backends.vllm.VLLMBackendneeds[vllm](Linux/CUDA only).
API backends. As of late 2025 every modern provider’s strict structured-outputs mode is real token-level masking inside the provider’s runtime — not post-validation. The honest split is “where the masking runs”, not “logit vs schema”:
citeformer.backends.openai.OpenAIBackendneeds[openai]. Strict JSON schema → token-level masking inside OpenAI ongpt-4o-2024-08-06+and successors.citeformer.backends.anthropic.AnthropicBackendneeds[anthropic]. Adapter over Anthropic’s native Citations API — provider enforces that every cite references a supplied document.citeformer.backends.gemini.GeminiBackendneeds[gemini]. Constrained generation viaresponse_schema(OpenAPI-subset).citeformer.backends.mistral.MistralBackendneeds[mistral]. Strict JSON schema (mistral-large-2411+).citeformer.backends.openrouter.OpenRouterBackendneeds[openrouter]. Multi-provider routing on the OpenAI wire format withprovider.require_parametersso requests refuse to land on upstreams that can’t honour strict mode.citeformer.backends.fireworks.FireworksBackendneeds[fireworks]. True logit-tier on a hosted API — Fireworks’s native GBNF grammar mode accepts citeformer’scite-idrule unchanged, so the same constraint that masks logits insideHFBackendruns inside the Fireworks runtime.citeformer.backends.together.TogetherBackendneeds[together]. Strictjson_schemaconstrained decoding on Together’s open-weight upstream models (Llama, Qwen, DeepSeek, …).
Per-provider tier discussion lives in
docs/reference/architecture.md.
Submodules¶
citeformer.backends.llamacppciteformer.backends.baseciteformer.backends.mockciteformer.backends.geminiciteformer.backends.openrouterciteformer.backends.vllmciteformer.backends.mistralciteformer.backends.anthropicciteformer.backends.togetherciteformer.backends.openaiciteformer.backends.hfciteformer.backends.fireworks