Assistant · On every page · In rNitro

chopsticksAI

A general-purpose assistant that answers anything you ask — and the in-house expert on every Chopsticks app. chopsticksAI v1.0, reachable from any page here and inside rNitro's Chat tab.

Version chopsticksAI v1.0 Drafts with NVIDIA Nemotron 3 Ultra 550B Reviewed by Google Gemma 4 26B (a4b) Context 48K tokens Limit 775K tokens · 3-hour cooldown Knowledge base 60 topics · 2,285 terms Your key never needed

macOS app

Native chat window with the same live model and offline knowledge-base fallback. No Xcode required to install.

curl -fsSL https://chopstickshq.com/chopsticks-ai/install-chopsticks-ai.sh | bash

How it answers

chopsticksAI v1.0

chopsticksAI v1.0 is the assistant built and run by Chopsticks HQ. Every answer is generated on request and grounded in the knowledge base below, so product details come from real reference material rather than guesswork.

Two models collaborate on each answer. NVIDIA Nemotron 3 Ultra (550B, 55B active) writes the draft, then Google Gemma 4 26B (26B, 4B active) reviews and tightens it — correcting wording and cutting repetition without adding facts the draft did not contain.

Generated files and code skip the review pass, so they are never silently rewritten. If the first model is rate-limited the next takes over, and if the review fails the draft is sent as-is. Every model in the chain is a zero-cost tier, so the second opinion adds quality rather than expense.

Requests run through a Chopsticks HQ endpoint that holds the API credential server-side. Nothing is embedded in this page or in the rNitro binary, and you never supply a key of your own.

Why it doesn't invent version numbers

A language model asked "what version is rNitro?" will happily invent a plausible answer. chopsticksAI doesn't, because the model never answers from memory alone.

Each question is first scored against a knowledge base of 60 curated topics, expanded into 2,285 matchable terms covering synonyms, plurals and common misspellings. The best-matching entries are injected into the prompt as authoritative reference material, and the model is instructed to prefer them over its own assumptions.

Matching is done on word boundaries rather than raw substrings — so "unzip the file" doesn't trigger the ZIP download topic, and "hardcore gaming" doesn't trigger the CPU-cores topic.

How a question travels

you ask a question ↓ retrieval score against 2,285 terms, take the 6 best topics ↓ grounding inject those facts as reference material ↓ search free sources consulted when the question needs current facts ↓ draft Nemotron 3 Ultra 550B writes the answer ↓ review Gemma 4 26B tightens and corrects it ↓ answer conversational, with real version numbers if the proxy is unreachable ↓ fallback the same knowledge base answers offline, on-device

Built in

No API key

You never paste a credential. The key lives as a server secret, so it is never shipped to a browser or extractable from the app.

Works offline

If the model can't be reached, the on-device knowledge base answers instead of erroring. The assistant degrades rather than dies.

Searches the web

When a question needs current facts, it consults free sources before answering and cites what it used. Product questions skip the lookup — those come from the knowledge base.

Nothing stored

No accounts, no chat logs, no analytics on what you ask. Consistent with every other Chopsticks app.

Budgeted

A 775,000-token allowance with a 3-hour cooldown, plus per-visitor rate limiting, so a traffic spike can't exhaust the quota.

All the files

The complete source. The knowledge base is written once in kb.source.json and compiled by build-kb.py into three targets — web, server and Swift — so the website and the app can never drift apart.

FileWhat it isSize
kb.source.json The knowledge base. 60 curated topics, hand-written. The one file to edit. 42 KB view
build-kb.py The compiler. Expands terms into synonyms, plurals and typos, then emits all three targets. Fails the build on ambiguous keywords. 10 KB view
chopsticks-ai.server.js The proxy. Holds the credential, retrieves grounding, calls the model, enforces the budget and rate limits. 11 KB view
chopsticks-ai.web.js The web client. The widget on every page, plus the offline engine used when the proxy is unreachable. 11 KB view
chopsticks-ai-kb.js Generated knowledge base for the browser. 63 KB view
ChopsticksAIEngine.swift The Swift engine. Scoring identical to the JavaScript one, shipped inside rNitro. 3 KB view
ChopsticksAIKB.swift Generated knowledge base for the app. 74 KB view
fixtures.json The tests. 69 question/answer pairs, including questions that must not match. 5 KB view
run-fixtures.mjs Test runner. Both engines must resolve all 69 identically. 2 KB view

Try it

Tap the button in the bottom-left corner of any page on this site. Inside rNitro, open the menu bar icon and pick the Chat tab — chopsticksAI is the default provider and needs no setup.