Search / NLP
Multilingual Sanctions Search
A multilingual screening reference that keeps normalization, retrieval and human review boundaries explicit.
Language handling
EN / RU / UK
normalized names, aliases and mixed-script inputs
Retrieval
Lexical + vector
exact and fuzzy candidates with optional semantic escalation
Reference suite
1,739 passed
with two optional FAISS checks skipped in the recorded local run
System view
Request-to-decision pathProblem
Sanctions screening has to reconcile spelling variants, aliases, initials, identifiers and multiple scripts. A plausible similarity score is not enough: the result must retain the signals and source record needed for human review.
Engineering approach
- Validate the request, detect language and normalize Unicode, names and entity fields.
- Retrieve exact, phrase and n-gram candidates before optional semantic escalation.
- Fuse candidates, resolve entities and return reviewable supporting signals.
- Fail closed when the screening index is unavailable or incomplete.
What the evidence supports
The public repository is an engineering reference and screening aid. The recorded local reference suite checks implementation behavior; it does not establish search precision, recall or legal suitability.