text-provenance

text-provenance

When to Use

• RAG citation highlighting — show which source document a generated text came from
• Contract playbook matching — find which standard clause a contract clause derives from
• Plagiarism detection
• Source attribution for AI-generated legal text
• Any text provenance task where you need to trace text back to its origin

How It Works

Core Approach

Lightweight text similarity metrics — no embeddings or API calls at runtime. Fast, deterministic string matching that works in-browser or server-side.

Comparison Methods

• Surface similarity — character-level comparison
• N-gram overlap — phrase-level matching
• Fingerprint matching — exact phrase detection

Usage

import { findProvenance } from 'text-provenance';

const sources = await findProvenance(
  "the quick brown fox jumps over the lazy dog",
  corpusDocuments
);
// Returns ranked list of potential sources with confidence scores

Edge Cases

Works where embeddings fail:

• Short text snippets
• Exact phrase matching
• High-precision attribution tasks
• Privacy-sensitive contexts (no data leaves the machine)

Audience and Work Shape

Audience: developers and lawyers building RAG-citation, contract-derivation, or plagiarism-detection workflows. The output is a candidate ranking, not a finding.

Work shape: Pattern-Matched Review. Lexical similarity is the matching function; the user decides what counts as a match.

Scope and Legal Use

This skill provides legal support, not legal advice. The output is a ranked list of candidate sources with similarity scores — never an attribution conclusion, never a plagiarism finding, never a contract-derivation determination of legal effect.

Privilege and confidentiality. Runs client-side with no network calls. No text leaves the user's machine unless the calling application chooses to transmit it. Skill itself does not create new privilege exposure.

Accountability. A qualified lawyer must review and accept any output before relying on it for an attribution, plagiarism, or contract-derivation decision. The similarity score is a signal, not a verdict.