Provider catalog¶

A Provider is a concrete implementation of a Capability — registered under latence.providers and resolved by its entry-point name. The framework never names a model; it names Capabilities and lets Providers fulfil them. Every Provider declares a ProviderProfile with its compute, footprint, and verified weights-and-code license — the columns below come straight from those declared objects.

Reading the catalog

Compute — cpu runs anywhere, gpu needs CUDA (skipped-with-flag on a CPU-only host via the device seam), either follows the requested device.
License — the SPDX of the model weights (or the code, for a pure-code Provider). Reference Providers ship permissive weights only; restricted models are opt-in (ADR-0012).
Endpoint Providers are thin HTTP clients (any OpenAI-compatible endpoint) — the served model lives behind the endpoint, so they carry no local weights.

Built-in reference Providers (ship in `latence-core`)¶

Every Stage has a deterministic, dependency-light reference Provider that runs on a laptop CPU, offline (ADR-0007). These ship inside core so the demo runs end-to-end with no extra install, and most Stages ship a second CPU reference Provider so the seam is exercised by real alternatives (ADR-0035).

Capability	Provider	Technique	Compute
Source	`source.local_folder`	Filesystem / object-store listing + fetch	cpu
Chunk	`chunk.markdown`, `chunk.sentence_window`, `chunk.page`	Offset-preserving markdown / sentence-window packing / page-preferred (one chunk per source page, 1024-token window otherwise)	cpu
Screening (intake)	`screening.intake_signature`	Signature + size intake screen	cpu
Screening (content)	`screening.content_keyword`, `screening.content_fuzzy`	Keyword / fuzzy injection detection	cpu
Entity Extraction	`entity.gazetteer`	Zero-shot gazetteer / regex	cpu
Relation Extraction	`relation.pattern`	Trigger-phrase patterns	cpu
Redaction	`redaction.hybrid_rule`	Regex library + gazetteer, sensitivity-keyed	cpu
Profiling	`profiling.statistical`, `profiling.lightweight`	Streaming corpus features	cpu
Disambiguation	`disambiguation.cascade`, `disambiguation.exact_surface`	5-strategy resolver / exact-surface blocking	cpu
Graph Assembly	`graph.canonical`, `graph.weighted`	Content-addressed nodes/edges + per-edge Evidence	cpu
Graph Completion	`graph_completion.reference`	Bounded 2-hop transitive/symmetric closure	cpu
Context Enrichment	`context.kg_header`	KG-header projection onto chunks	cpu
Embedding	`embedding.hashing`	Deterministic hashing embedder	cpu
Delta	`delta.affected_set`	Affected-set recompute + drift reconcile	cpu
Export	`export.jsonl_parquet`, `export.knowledge_graph`	JSONL/Parquet + TTL/GraphML files	cpu

Source¶

Connectors to source systems live outside core as their own Provider packages (ADR-0014) — the Source seam is what makes a real connector a plugin rather than a fork.

Provider	Package	Technique	License	Compute
`source.sharepoint`	`latence-source-sharepoint`	SharePoint / OneDrive via Microsoft Graph; content-hash Delta skips unchanged documents without downloading them; secrets env-only	Apache-2.0 (code), no third-party dependency	cpu

Parse¶

Provider	Package	Model / technique	License	Compute
`parser.plaintext`	`latence-parser-plaintext`	Passthrough plain-text / markdown	Apache-2.0 (code)	cpu
`parser.document`	`latence-parser-document`	PDF + text → markdown + page map	Apache-2.0 (code)	cpu
`parser.pdfplumber`	`latence-parser-pdfplumber`	Page-aware PDF text layer (pdfplumber)	MIT (code)	cpu
`parser.render`	`latence-parser-render`	Universal any-file render → OCR, license-clean (no PyMuPDF/AGPL)	Apache-2.0 (code)	either
`parser.lighton`	`latence-parser-lighton`	`lightonai/LightOnOCR-1B-1025` (1B VLM OCR)	Apache-2.0	gpu
`parser.glm`	`latence-parser-glm`	`zai-org/GLM-OCR` (0.9B VLM OCR)	MIT	gpu
`parser.lighton_vllm`	`latence-parser-lighton-vllm`	`lightonai/LightOnOCR-2-1B` served via vLLM	Apache-2.0	either (endpoint)
`parser.endpoint`	`latence-parser-endpoint`	PaddleOCR-VL / any vision endpoint	Apache-2.0 (client)	either (endpoint)

Entity & Relation Extraction¶

Provider	Package	Model / technique	License	Compute
`entity.gliner`	`latence-ner-gliner`	`urchade/gliner_multi-v2.1` (zero-shot NER)	Apache-2.0	either
`entity.endpoint`	`latence-ner-endpoint`	GLiNER-large / any endpoint	Apache-2.0 (client)	either (endpoint)
`relation.gliner_relex`	`latence-relation-gliner`	`knowledgator/gliner-relex-multi-v1.0` (fused NER+RE)	Apache-2.0	either
`relation.llm`	`latence-relation-llm`	Any OpenAI-compatible LLM (incl. local Ollama)	Apache-2.0 (client)	either (endpoint)
`fused_entity_relation.gliner2`	`latence-extract-gliner2`	`fastino/gliner2-multi-v1` (mdeberta-v3, fused)	Apache-2.0	either
`fused_entity_relation.gliner25`	`latence-gliner25`	`fastino/gliner2.5-multi-v1` (mDeBERTa-v3 boundary head, fused, 4096-token window)	Apache-2.0	either

Redaction (PII)¶

Provider	Package	Model / technique	License	Compute
`redaction.gliner_pii`	`latence-pii-gliner`	`urchade/gliner_multi_pii-v1` (zero-shot PII)	Apache-2.0	either
`redaction.gliner2`	`latence-pii-gliner2`	`fastino/GLiNER2-Guardrails-PII-Multi`	Apache-2.0	either
`redaction.gliner25`	`latence-gliner25`	`fastino/gliner2.5-multi-v1` — zero-shot PII on the GENERAL checkpoint (there is no GLiNER 2.5 PII checkpoint); recall UNMEASURED	Apache-2.0	either
`redaction.presidio`	`latence-pii-presidio`	Microsoft Presidio + spaCy `en_core_web_sm`	MIT	either

Disambiguation¶

Provider	Package	Model / technique	License	Compute
`disambiguation.embedding`	`latence-disambig-embedding`	Learned surface+context resolution (registry Embedder)	Apache-2.0 (code)	either
`disambiguation.glinker`	`latence-disambig-glinker`	`knowledgator/gliner-linker-large-v1.0` neural linker + audited resolver	Apache-2.0	gpu

Graph Completion & Schema Induction (optional Stages)¶

Provider	Package	Model / technique	License	Compute
`graph_completion.ultra`	`latence-linkpred-ultra`	Inductive ULTRA (`ultra_3g`) zero-shot scorer + PyKEEN fallback	MIT	either
`label_inducer.llm`	`latence-schema-inducer`	Any small OpenAI-compatible LLM induces each chunk's label schema (hosted, OpenRouter, or a LOCAL vLLM/Ollama box — `base_url` only)	Apache-2.0 (client)	either (endpoint)
`label_inducer.lexicon`	`latence-schema-inducer`	Deterministic zero-network tagger over an operator-declared candidate type space — the non-frontier per-chunk option (no model, no calls)	Apache-2.0	cpu

Embedders¶

Provider	Package	Model / technique	License	Compute
`embedding.sentence_transformers`	`latence-embedder-st`	`ibm-granite/granite-embedding-311m-multilingual-r2` (GPU 311m / CPU 97m)	Apache-2.0	either
`embedding.endpoint`	`latence-embedder-endpoint`	`multilingual-e5-large` / any embeddings endpoint	Apache-2.0 (client)	either (endpoint)

See Embedders for how the opt-in RAG-corpus vectors fit into Export, and ADR-0045 for why Granite r2 is the default.

Runners¶

Not Providers, but the other swap point: a Runner schedules the identical Stages on a substrate. Alongside the built-in local Runner, latence-runner-airflow runs the same Stages under Apache Airflow using Airflow's own state and scheduling (ADR-0003, ADR-0030).

Build your own¶

The catalog is meant to grow. To add a Provider — implement a Capability, register the entry point, declare a ProviderProfile, and pass conformance — follow Authoring a Provider, then compare it against the incumbents with a bake-off.