🎲 A random word generator designed to destabilize standard LLM pattern and probability-based thinking — fully offline.
All LLMs run on patterns and probabilities — they predict what should come next, but never break from their training data. Even when asked to "be random," they can't truly leave that internal loop — they're still inside the fishbowl. 🐠
The Rabbit Response Team breaks that loop. It picks truly random words from a bundled offline word bank and injects them directly into the AI's prompt. Each injected word must be used naturally in the model's response — forcing the AI to adapt, improvise, and escape pattern lock.
No network calls. No accounts. No API keys. Just words.
- 🎲 Three Randomness Modes —
random,double-pass(anchor + associations), andcontextual(keywords pulled from your latest message). - 🌐 English & Russian — bundled word banks for both languages, plus an Auto-detect mode that picks per turn based on script.
- 🔁 Synonym Scanner — detects overused words in recent chat history and surfaces fresh alternatives.
- 🪶 Prompt Injection — words are spliced into the prompt at a configurable depth (system prompt vs. user-message boundary).
- 🚫 Blacklist & History — keep specific words out and avoid repeats within a configurable window.
- 🔧 Custom Prompt Templates — fully editable, with reset-to-default buttons.
- 📴 Fully Offline — runs entirely inside SillyTavern with zero runtime dependencies.
- Go to Extensions → Extension Settings
- Find 🐇 Rabbit Response Team
- Toggle the feature(s) you want (Random Words, Synonyms, or both)
- English — uses the EN word bank only.
- Russian — uses the RU word bank only.
- Auto-detect — picks per turn based on the script of your latest message.
- Random — filtered sample (length, part of speech, blacklist, history).
- Double-pass — picks one anchor word with associations, then themes the remaining slots around it.
- Contextual — extracts keywords from your latest message and pulls in semantically related words.
Click 🎲 Test Random Words to preview what the extension will inject. Each test gives a completely different set of words.
The injected words will appear invisibly in your prompt's backend, and the AI will weave them naturally into its response.
git clone <this repo> into SillyTavern/public/extensions/third-party/
cd Rabbit-Response-Team
npm install # devDependencies only (wordnet-db for the build pipeline)
npm test # runs the node:test unit suite
npm run build # rebuilds assets/{en,ru}/*.json from scripts/raw/The extension runs directly in SillyTavern's client-side JavaScript environment — no bundler, no TypeScript.
Bundled word banks and synonym maps live under assets/{en,ru}/. To regenerate them from upstream sources (Datamuse wordfreq, wordnet-db, Badestrand Russian Dictionary, YARN), see scripts/raw/README.md for download URLs, expected filenames, and SHA256 pins.
It's a small experiment, but one I really believe will work! Let me know if you like it.
