arche-core

Know what's real.

The open engine for messy, multilingual data. Find the entities, resolve who they actually are, protect them under the law that applies, and sign every decision.

arche tells you and your agent what — and who — your data is actually talking about. It finds the entities in any document or system, standardises them for the names and addresses people actually use, and works out which real-world thing each one refers to. Along the way it keeps disagreement between sources instead of erasing it, protects everything under the law that applies, and signs every decision.

Two records read Damini Ogulu and Burna Boy. Three clinic registers read Fatima Abdullahi, Fatuma Abdullahi, F. Abdulahi. Two sanctioned men share the name Khalid Mehmood, the same country and the same programme, and have different fathers and different national IDs. Getting those three cases right — two splits to heal, one merge to refuse — is the whole job.

Calibrated on the world's hardest identity data, and built for how the world actually writes names and addresses everywhere. Africa is where the engine was made good, not the limit of where it works: the launch statute packs cover Nigeria, Kenya, South Africa and Ghana alongside GDPR and HIPAA Safe Harbor.

pip install arche-core

Optional extras add heavier capabilities:

pip install "arche-core[doc]"       # PDF, DOCX, PPTX, XLSX, HTML parsing
pip install "arche-core[detect]"    # GLiNER2-PII soft-PII detection
pip install "arche-core[presidio]"  # Microsoft Presidio integration
pip install "arche-core[resolve]"   # Splink + DuckDB resolution support

from arche import Pipeline

pipeline = Pipeline(jurisdiction="NG")
result = pipeline.process(
    "Fatima Abdullahi, NIN 12345678901, BVN 22100987654."
)

print(result.redacted_text)
# NAME_... NAME_..., NIN [NIN], BVN [BVN].

for detection in result.detections:
    print(
        detection.category,
        detection.sensitivity_tier.value,
        detection.regulatory_citation,
    )

Same pipeline shape works for the four launch jurisdictions:

How Huduma Namba is detected, because the distinction matters. A Huduma Namba has no check digit and shares its length range with NHIF numbers and the Kenyan National ID, so a bare 8-digit string carries no evidence of which of the three it is. Detection is therefore cue-anchored: PII-2-HUDUMA is emitted where the text says "Huduma", and a bare number is reported as whatever the bare-digit patterns make of it. That is the same evidence a person uses, and it is why this pattern carries 0.88 confidence where the bare-digit ones sit at 0.40–0.45. Until v0.3.0a2 there was no pattern at all and Huduma numbers were reported as PII-2-NHIF — mislabelled rather than missed.

Pipeline.process(...) finds identity signals and PII spans in text. With arche-core[doc], Pipeline.process_file(...) parses supported document files and runs the same detection pipeline over extracted text.

Every detection can carry a category, sensitivity tier, regulatory citation, and policy action. The policy layer maps detections to one of six actions:

Resolution depends on clean, protected signals. Arche helps prepare those signals by normalizing phones, identifying IDs, detecting names, parsing address fragments, and producing deterministic tokens that can become privacy-preserving join keys.

For current resolution workflows, install:

pip install "arche-core[resolve]"

from collections import Counter
from arche import Pipeline

pipeline = Pipeline(jurisdiction="ZA")
result = pipeline.process_file("dsar_response.pdf")

print(Counter(d.category for d in result.detections))
print(Counter(o.action for o in result.policy_outcomes))
print(result.redacted_text)

Result is a plain dataclass with the fields document_hash, detections, addresses, policy_outcomes, redacted_text, audit_log, metadata. There is no summary() helper — count over detections / policy_outcomes instead. On Pipeline(jurisdiction="ZA").process("ID 8001015009087, phone 082 555 1234.") the two Counter lines print:

Counter({'PII-2-NATIONAL_ID': 1, 'PII-3-PHONE': 1})
Counter({'mask': 1, 'tokenize': 1})

from arche import match

score = match("Mamadou Diallo", "Muhammad Jallow", jurisdiction="NG")
print(score.decision, score.score)
# match 0.8865

Use this when you need culturally aware name matching before or after PII detection.

• Docs: https://unpatterned-labs.github.io/arche/
• Quick start: docs-site/docs/getting-started/quickstart.md
• API reference: docs-site/docs/api/index.md
• Examples: examples/

The framework is Apache-2.0. Dataset licensing is documented separately in the dataset cards and repository licensing files.