RhetoriLex

Free local academic paraphrase skill and evidence-safe writing workbench for researchers, students, editors, and AI agents.

Open the workbench · Explore the catalog · Bahasa Indonesia · Agent Skill · Research · Provenance

RhetoriLex helps writers paraphrase, audit, and revise academic or technical prose from communicative purpose and evidence, not from how academic it sounds. Version 0.2.0 contains 96 independently authored patterns across 24 rhetorical functions, with explicit claim strength, evidence requirements, causal-design guards, risk, named slots, and provenance.

Version 0.2.0 also indexes 144 original writing skills: 64 academic, 66 scientific, and 14 paraphrasing. The bilingual site publishes 62 localized routes: 31 English and 31 Indonesian, including an installable local-first paraphrase workbench.

It does not write evidence for you. It does not invent citations. It does not turn an association into a cause. Final submission checks remain the writer's responsibility.

Why it is different

• Evidence before polish. Every pattern declares the minimum evidence and strongest compatible claim.
• Meaning-preserving. The Agent Skill protects citations, numbers, units, populations, negation, uncertainty, and causal status during rewriting.
• Clean-room catalog. Released wording is original editorial work with source_reuse: false; restricted source inventories are excluded.
• Offline core. Search, filters, validation, builds, and the Skill helper use only Python's standard library.
• One source of truth. Canonical JSONL deterministically produces JSON, CSV, SQLite, Markdown, package resources, Skill assets, and checksums.
• Installable web workbench. The PWA can build safe prompts without AI, or call a user-supplied OpenAI-compatible or Gemini-compatible endpoint with a browser-side 3-call daily guard.
• Five useful surfaces. Use the Python API/CLI, portable Agent Skill/plugin, static browser explorer, local-first paraphrase workbench, or generated data artifacts.

Try it

Open the static workbench at https://rezaprama.github.io/RhetoriLex/en/paraphrase-workbench/. Use prompt-only mode for no remote AI call, or enter your own local/provider endpoint and token. The token is handled in the browser and stored only if you choose to remember settings.

git clone https://github.com/rezaprama/RhetoriLex.git
cd RhetoriLex
python -m pip install -e .
rhetorilex search "cautious interpretation" --stage discussion --limit 3

Machine-readable output and hard evidence filters:

rhetorilex --json search "observational result" \
  --evidence observational \
  --max-claim-strength bounded \
  --risk low

Inspect one pattern or the full contract health:

rhetorilex explain RLX-QUA-001
rhetorilex inspect --validate
rhetorilex taxonomy

Python API:

from rhetorilex import Catalog

catalog = Catalog.load()
results = catalog.search(
    "state a limitation",
    stage="discussion",
    max_claim_strength="bounded",
    limit=3,
)

for result in results:
    print(result.entry.template, result.entry.evidence_requirement)

Install the Agent Skill

For one repository, copy skills/rhetorilex to .agents/skills/rhetorilex. For personal use across projects, copy it to ~/.codex/skills/rhetorilex.

The Skill can route requests such as:

• “Give three cautious ways to interpret this observational result.”
• “Rewrite this paragraph without changing citations, values, or claim strength.”
• “Compare options for stating a bounded research gap.”
• “Audit this sentence for causal overclaiming.”

It intentionally refuses citation fabrication, plagiarism disguise, synonym spinning, detector evasion, and unsupported claim upgrades. A release-ready plugin archive can be built from .codex-plugin/plugin.json.

Catalog model

Each canonical entry includes:

See data/taxonomy/taxonomy.v1.json, data/contracts/evidence-claim.v1.json, and docs/data-model.md.

Research and discovery

The model is informed by research on genre and rhetorical moves (10.1017/CBO9781139524827), stance and reader interaction (10.1177/1461445605050365), and corpus phraseology and disciplinary variation (10.1016/j.jeap.2019.01.003). These sources orient original design; their prose, examples, and inventories are not incorporated. Full research provenance is in PROVENANCE.md.

Public positioning is also measured, not guessed. The discovery study and GitHub landscape record queries, metrics, timestamps, interpretation limits, and structured evidence under data/discovery. GitHub repository counts are supply/index signals—not web-search volume, user demand, adoption, or quality. Direct Google Trends measurement was unavailable, so its recorded values are explicit null; no proxy volume was invented.

Build and verify

python scripts/validate_data.py
python scripts/build_data.py
python -m unittest discover -s tests -v
python scripts/package_plugin.py --output dist/rhetorilex-plugin.zip

The test suite checks schemas, provenance, slot equality, evidence requirements, causal guards, deterministic builds, retrieval behavior, frozen benchmark routing, CLI behavior, and contract safety. Continuous integration repeats builds and compares checksums.

Architecture

data/canonical/catalog.v1.jsonl
          |
          +--> data/dist/*                    portable data artifacts
          +--> src/rhetorilex/resources/*     Python package
          +--> skills/rhetorilex/assets/*     offline Agent Skill
          +--> docs/data/phrases.json          assembled Pages artifact

taxonomy + JSON Schema + evidence contract
          |
          +--> validator --> tests --> deterministic release gate

The browser explorer and paraphrase workbench are static HTML, CSS, and JavaScript. They have no account, backend, CDN dependency, telemetry, or RhetoriLex manuscript database. Endpoint calls go directly from the user's browser to the endpoint they enter. A future shared-key hosted service would need a backend with server-held secrets, quota storage, abuse protection, and manuscript logging disabled by default.

Source and rights boundary

The private workbook supplied for this project carries an academic-use and no-redistribution notice. It was inspected locally only for non-expressive aggregate structure and quality signals. It is ignored by Git, its cells and labels are absent from this repository, every migration candidate remains unresolved, and zero source-derived items were promoted.

Manchester Academic Phrasebank is citation/conceptual context only; its phrase inventory was not imported or rewritten. BAWE is excluded. The Elsevier OA CC-BY Corpus v3 is registered only as a possible future aggregate-validation source and was not ingested. Full decisions, source links, and review rules are in PROVENANCE.md, data/sources/source-registry.yaml, and reports/xlsx-audit.md.

Contributing

Contributions are welcome when wording is independently authored and evidence-safe. Start with CONTRIBUTING.md. New patterns require provenance, named slots, taxonomy compatibility, evidence metadata, originality review, and tests. Never attach a restricted workbook, source inventory, manuscript, or corpus excerpt to an issue or pull request.

Security, confidential-draft, provenance, and rights concerns belong in the private process described by SECURITY.md. Project decisions are recorded in GOVERNANCE.md and IMPLEMENTATION_LOG.md.

Project and creator

RhetoriLex is created and maintained by Reza Prama Arviandi. Release history is in CHANGELOG.md; the public roadmap is in ROADMAP.md.

License and citation

Software, configuration, tests, and Agent Skill code are Apache-2.0. Original data, reports, benchmarks, assets, and documentation are CC BY 4.0. Machine-readable path mapping lives in REUSE.toml; third-party status lives in THIRD_PARTY_NOTICES.md.

Citation metadata for version 0.2.0 is available in CITATION.cff. Licensing documents state project policy, not legal advice.