anno

Extract named entities, relations, coreference chains, and PII from unstructured text. Fixed entity types (PER/ORG/LOC/MISC) or zero-shot custom labels.

Dual-licensed under MIT or Apache-2.0. MSRV: 1.85.

Quickstart

[dependencies]
anno = "0.3.9"

let entities = anno::extract("Sophie Wilson designed the ARM processor.")?;
for e in &entities {
    println!("{} [{}] ({},{}) {:.2}", e.text, e.entity_type, e.start(), e.end(), e.confidence);
}
// Sophie Wilson [PER] (0,13) 0.95
// ARM [ORG] (27,30) 0.90
# Ok::<(), anno::Error>(())

Filter results with prelude:

use anno::prelude::*;

let people: Vec<_> = entities.of_type(&EntityType::Person).collect();
let confident: Vec<_> = entities.above_confidence(0.8).collect();

For backend control, construct a model directly:

use anno::{Model, StackedNER};

let m = StackedNER::default();
let ents = m.extract_entities("Sophie Wilson designed the ARM processor.", None)?;
# Ok::<(), anno::Error>(())

StackedNER::default() selects the best available backend at runtime: BERT or NuNER (if onnx enabled and models cached), then GLiNER, falling back to heuristic + pattern extraction. Set ANNO_NO_DOWNLOADS=1 or HF_HUB_OFFLINE=1 to force cached-only behavior.

Zero-shot custom types via GLiNER:

use anno::GLiNEROnnx;

let m = GLiNEROnnx::new("onnx-community/gliner_small-v2.1")?;
let ents = m.extract("Aspirin treats headaches.", &["drug", "symptom"], 0.5)?;
for e in &ents {
    println!("{}: {}", e.entity_type, e.text);
}
// drug: Aspirin
// symptom: headaches
# Ok::<(), anno::Error>(())

Custom backends

AnyModel wraps a closure into a Model, bypassing the sealed trait when you need to plug in an external NER system:

use anno::{AnyModel, Entity, EntityType, Language, Model, Result};

let model = AnyModel::new(
    "my-ner",
    "REST API wrapper",
    vec![EntityType::Person, EntityType::Organization],
    |text: &str, _lang: Option<Language>| -> Result<Vec<Entity>> {
        Ok(vec![]) // call your backend here
    },
);
let ents = model.extract_entities("test", None)?;
# Ok::<(), anno::Error>(())

What it does

Named entity recognition. Spans (start, end, type, confidence) with character offsets (Unicode scalar values, not bytes). Fixed taxonomies (PER/ORG/LOC/MISC) or caller-defined labels for zero-shot extraction [1, 2].

Coreference resolution. Group mentions into clusters tracking the same referent. Rule-based sieves (SimpleCorefResolver), neural (FCoref, 78.5 F1 on CoNLL-2012 [3]), and mention-ranking (MentionRankingCoref).

Structured patterns. Dates, monetary amounts, emails, URLs, phone numbers via deterministic regex grammars.

Relation extraction. (head, relation, tail) triples via RelationCapable backends (gliner2, tplinker). Other backends produce co-occurrence edges for graph export.

PII detection. Classify NER entities as PII and scan for structured patterns (SSN, credit card, IBAN, email, phone). Redact or pseudonymize in one call:

use anno::{pii, Model, StackedNER};

let text = "John Smith's SSN is 123-45-6789.";
let m = StackedNER::default();
let ents = m.extract_entities(text, None)?;

let mut pii_ents: Vec<_> = ents.iter().filter_map(pii::classify_entity).collect();
pii_ents.extend(pii::scan_patterns(text));
let redacted = pii::redact(text, &pii_ents);
// "[REDACTED]'s SSN is [REDACTED]."
# Ok::<(), anno::Error>(())

Export. Brat standoff, CoNLL BIO tags, JSONL, N-Triples, JSON-LD, and graph CSV via pure functions in anno::export.

Backends

See BACKENDS.md for details, default models, and WIP backends.

ML backends are feature-gated (onnx or candle). Weights download from HuggingFace on first use.

Feature flags

CLI

cargo install --git https://github.com/arclabs561/anno --package anno-cli --bin anno --features "onnx"

anno extract --text "Lynn Conway worked at IBM and Xerox PARC in California."
# PER:1 "Lynn Conway"
# ORG:2 "IBM" "Xerox PARC"
# LOC:1 "California"

anno extract --model gliner --extract-types "DRUG,SYMPTOM" \
  --text "Aspirin can treat headaches and reduce fever."
# drug:1 "Aspirin" symptom:2 "headaches" "fever"

anno debug --coref -t "Sophie Wilson designed the ARM. She revolutionized mobile computing."
# Coreference: "Sophie Wilson" -> "She"

JSON output with --format json. Batch processing with anno batch. Graph export (N-Triples, JSON-LD, CSV) with anno export --features graph.

Coreference

FCoref requires a one-time model export: uv run scripts/export_fcoref.py (from a repo clone).

RAG preprocessing (rag::resolve_for_rag(), rag::preprocess()): rewrites pronouns for self-contained chunks after splitting. Always available (no feature flag required).

Scope

Inference-time extraction. Training pipelines are out of scope -- use upstream frameworks and export ONNX weights.

Troubleshooting

• ONNX linking errors: use default-features = false for builds without C++, or check ORT_DYLIB_PATH.
• Model downloads: set HF_HUB_OFFLINE=1 for cached-only mode behind firewalls.
• Feature errors: most backends are gated behind onnx or candle.
• Offset mismatches: all spans use character offsets, not byte offsets. See CONTRACT.md.

Documentation

• QUICKSTART -- getting started
• CONTRACT -- offset semantics, scope
• BACKENDS -- backend details, feature flags
• ARCHITECTURE -- crate layout
• REFERENCES -- full bibliography
• API docs

References

[1] Grishman & Sundheim, COLING 1996. [2] Tjong Kim Sang & De Meulder, CoNLL 2003. [3] Otmazgin et al., AACL 2022 (F-COREF). [4] Jurafsky & Martin, SLP3 2024. [5] Zaratiana et al., NAACL 2024 (GLiNER). [6] Bogdanov et al., 2024 (NuNER). [7] Li et al., AAAI 2022 (W2NER). [8] Devlin et al., NAACL 2019 (BERT). [9] Lafferty et al., ICML 2001 (CRF). [10] Wang et al., COLING 2020 (TPLinker). [11] Zaratiana et al., 2025 (GLiNER2). [12] Rabiner, Proc. IEEE 1989 (HMM).

Full list: docs/REFERENCES.md. Citeable via CITATION.cff.

License

Dual-licensed under MIT or Apache-2.0.