What the engine does

noplag-engine is the open-source plagiarism detection engine that powers the hosted Noplag product. It is Apache-2.0 licensed, self-hostable, and built on one belief: auditable beats black-box. Trust in a plagiarism score shouldn't require taking a vendor's word for it.

Scope, honestly

The engine detects verbatim and near-verbatim copying: text that was pasted, lightly edited, reordered, or partially reworded but still shares literal character-level overlap with a source.

It does not detect:

• heavy paraphrase or summarization,
• cross-language plagiarism,
• AI-generated writing.

Those are on the public roadmap as separate layers (paraphrase, semantic, AI detection) — they are not in this codebase, and no configuration flag turns them on.

How it works

1. Chunking — documents are split into sentence-window chunks.
2. Fingerprinting — each chunk is reduced to winnowing fingerprints, stored as a GIN-indexed bigint[] in Postgres.
3. Retrieval (L1) — candidate source chunks are retrieved by fingerprint overlap.
4. Alignment (L0) — matches are refined to exact character ranges with a seed-and-extend aligner.

The result is a report with per-source similarity, per-passage character offsets on both the query and source side, and honest coverage accounting — if a check hits its time budget, it says so instead of silently truncating.

What's different from the hosted product

The open-source engine matches against your local corpus only — it ships with a small Wikipedia sample corpus and tools to bring your own. The hosted product additionally searches the live web and maintains a large multi-tier corpus.

Three commitments

• Auditable — Apache-2.0. Read it, fork it, self-host it.
• Honest — every match shows its source, and we don't claim accuracy we can't measure. See Benchmarks for exactly what the numbers mean.
• Yours — your documents are never used for training, and a self-hosted engine runs entirely on your own infrastructure.