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.

v0.3.0a1 · pre-beta · Apache-2.0

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.

detect

Find the entities and the identifying data in text and documents.

resolve

Work out which real-world thing each reference points at — and abstain when the evidence does not support a verdict.

protect

Apply the statute that governs the data, and cite the section it came from.

attest

Sign a decision together with the evidence and the representation that produced it.

The problem, in three records

Two catalogue rows read Damini Ogulu and Burna Boy. Three clinic registers around Kano read Fatima Abdullahi, Fatuma Abdullahi, F. Abdulahi. And two sanctioned men share the name Khalid Mehmood, the same country and the same programme, with different fathers and different national IDs.

Two of those are one entity wearing different names. One is two entities wearing the same name. Software that gets the first two right by loosening its matching gets the third one catastrophically wrong. Holding all three at once is the job.

Why use this library

• It abstains. When the evidence does not support a verdict, resolution returns review rather than guessing. An agent that flips two people has a much worse day than one that asks.
• Every decision cites its law. Detections carry a sensitivity tier and the specific statute section that classifies them — NDPA-2023, POPIA, Kenya DPA, Ghana DPA, GDPR, or HIPAA Safe Harbor.
• Every decision can be signed together with the exact representation that produced it, so the claim given this evidence and this representation, this was the decision is checkable by anyone.
• Your schema, not ours. One YAML declares your fields; arche generates the comparators, the masking, and the extraction contract your LLM fills.
• Bring any model. An LLM is a proposer, never the decider. It reads messy text into your declared fields — hallucinated fields become violations rather than values — and the engine grades its judgment against a deterministic oracle, counting review as an honest abstention rather than a miss. The integration surface is one callable. Bring your own LLM.
• Calibrated on the hardest identity data there is. Africa is where the engine was made good, not the limit of where it runs. The name equivalence and frequency data ships as inspectable files you can read and correct.
• Lightweight by default. Heavy ML, Presidio, Splink, and document parsing are opt-in extras.

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 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 - there is no summary() helper. Its fields are document_hash, detections, addresses, policy_outcomes, redacted_text, audit_log, and metadata; counting over detections / policy_outcomes is how you get a per-category and per-action rollup. On the equivalent inline call Pipeline(jurisdiction="ZA").process("ID 8001015009087, phone 082 555 1234.") those two lines print:

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

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 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.

What ships today

Verb	What arche does today
detect	Finds entities, PII, and identity signals in text and supported document files
resolve	Scores a pair or a whole table, returns same_entity / review / different, and abstains rather than guessing
protect	Applies jurisdiction-aware policy actions and emits audit-ready output with the statute section cited
attest	Signs a decision with its evidence and the representation that produced it

Next steps

• Getting started
• Match African names
• Extract from an invoice
• Nigerian fintech KYC cookbook
• Introducing arche v0.2
• Pipeline API reference

Licence

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