arche-core¶

The identity data engine for Africa.

Arche finds identifying data, helps protect it according to the right jurisdiction, and prepares it for privacy-preserving resolution.

The public arche-core package focuses on three connected jobs:

Detect identifying data in African text and documents.
Protect it with jurisdiction-aware masking, tokenization, generalization, dropping, retention, and audit actions.
Resolve more safely by producing normalized, policy-aware signals such as tokenized IDs, names, phones, and addresses.

Today, Detect and Protect are the lead product surface. Resolution support is intentionally narrow: name matching, tokenized identifiers, and optional Splink-backed workflows for larger linkage tasks.

Why use this library¶

Simple. One Pipeline.process(...) call runs detection, policy, redaction, and audit output.
African-first. Launch support covers Nigeria, Kenya, South Africa, and Ghana, with wider African identifier, phone, name, and address support.
Statute-aware. Detections are grounded in NDPA-2023, POPIA, Kenya DPA, or Ghana DPA policy files.
Lightweight by default. Heavy ML, Presidio, Splink, and document parsing dependencies are opt-in extras.
Useful for review workflows. You can scan text, PDFs, DOCX files, invoices, DSAR responses, leaked documents, KYC records, and review extracts without building a separate compliance layer first.

Installation¶

pip install arche-core

For document parsing:

pip install "arche-core[doc]"

Optional extras include:

Extra	Adds
`arche-core[doc]`	docling-backed PDF, DOCX, PPTX, XLSX, and HTML parsing
`arche-core[doc-ocr]`	OCR support for scanned documents
`arche-core[detect]`	GLiNER2-PII soft-PII detection
`arche-core[presidio]`	Microsoft Presidio integration
`arche-core[resolve]`	Splink and DuckDB entity resolution support

What does it do?¶

Given text or a supported document file, arche returns:

the detected PII spans
their taxonomy category
a sensitivity tier
the statute citation used by the loaded jurisdiction
the policy action applied
redacted text
audit records suitable for later review

These outputs are useful for redaction today and for safer record linkage later: tokenized IDs, normalized phones, detected names, and parsed address fragments can become privacy-preserving join signals.

Supported launch jurisdictions:

Jurisdiction	Policy loaded
`NG`	NDPA-2023
`ZA`	POPIA
`KE`	Kenya DPA
`GH`	Ghana DPA

Example: detect PII in text¶

from arche import Pipeline

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

print(result.redacted_text)

Example output:

NAME_... NAME_..., NIN [NIN], BVN [BVN].

You can inspect the detections directly:

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

Example output:

Category	Tier	Citation
`PII-2-NIN`	`high`	`NDPA-2023 s.30, NIMC Act s.27`
`PII-2-BVN`	`high`	`NDPA-2023 s.30, CBN BVN policy 2014`
`PII-1-NAME`	`moderate`	`NDPA-2023 s.30`

Example: scan a document¶

With arche-core[doc] installed, use the same pipeline on files:

from arche import Pipeline

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

print(result.summary())
print(result.redacted_text)

process_file(...) delegates parsing to the document substrate, then sends the extracted text through the same detection and policy pipeline.

What can it detect?¶

Area	Current coverage
Government IDs	NG NIN, BVN, TIN, RC, PVC, drivers licence; KE National ID, Huduma, KRA PIN, NHIF; ZA ID, tax, passport; GH Ghana Card, SSNIT, TIN; plus wider African ID patterns
Names and local NER	African name lexicon and equivalence data, with optional GLiNER soft-PII detection
Phones	libphonenumber-backed E.164 normalization across African networks
Addresses	Nigeria and South Africa parser MVP
Digital identifiers	DIDs, Bitcoin addresses, Ethereum addresses
Network identifiers	IPv4 and IPv6 detection with private and special-range flags

Matching names¶

from arche.match import match

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

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

Detect, Protect, Resolve¶

Step	What Arche does today
Detect	Finds PII and identity signals in text and supported document files
Protect	Applies jurisdiction-aware policy actions and emits audit-ready output
Resolve	Prepares normalized, protected signals for matching and linkage workflows

Next steps¶

Licence¶

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