Most AI tools will confidently invent an answer when they don't actually know one. A made-up citation, a plausible-sounding statistic, a function that doesn't exist delivered with the same confidence as the truth. For anything that matters, that confidence is the problem.
The I Don't Know Project takes the opposite stance. It answers only from a specific library of sources you give it. Every answer shows the exact source behind it. And when the answer isn't in those sources, it plainly says "I don't know" rather than guessing.
It's not trying to know everything. It's trying to be trustworthy about what it knows and honest about what it doesn't.
- Answers only from loaded sources. No open web, no training-data guesswork. If it's not in the library, it isn't in the answer.
- Shows its receipts. Every answer is tied to the source it came from, so you can verify it yourself in one click.
- Says "I don't know." When the sources don't cover your question, it says so — and lets you flag the gap so the library can grow where people actually need it.
- Grades its own confidence. Answers are marked high / medium / low confidence based on how well the sources support them, so a thin answer never masquerades as a sure one.
- Resolves conflicts. When sources disagree, it follows a defined rule rather than picking at random.
- Built-in honesty guardrails. The engine is designed to resist prompt-injection and to fail closed — when in doubt, it abstains.
It ships with a small demo library so you can see it work immediately. Point it at your own sources — documentation, policies, guidelines, a body of texts, anything — and it becomes a trustworthy answer engine for that domain.
This is open source under the MIT license. Deploy it, fork it, rebrand it, build on it for your own documentation, your own knowledge base, your own community. That's the point. The honesty engine is the gift; what you point it at is up to you.
A deliberately simple, modern, low-cost stack easy to read, fork, and self-host.
- Frontend: React + Vite (single-page app)
- Backend: Cloudflare Pages Functions (serverless)
- Database: Cloudflare D1 (SQLite at the edge)
- AI: Anthropic Claude (model is configurable)
- Hosting: Cloudflare Pages — runs comfortably on free tiers
Beyond the core honesty engine, the project includes a full, working product around it:
| Area | What's there |
|---|---|
| Answers | Source-grounded answers, inline citations, confidence grading, conflict resolution |
| Sources | Curated source library, public source submissions, admin moderation queue |
| Accounts | Secure sign-up / login, saved conversation history per user |
| Community | A public "gap board" where people request sources and flag answers, votes included |
| Sharing | Share the project or share a single answer as its own public, cited page |
| Transparency | A live public changelog driven straight from the project's commit history |
A full setup guide is on the way. In short: it's a standard Cloudflare Pages + D1 project.
# Install dependencies
npm install
# Run locally
npm run dev
# Build for production
npm run buildDeployment uses Cloudflare Pages with a D1 database binding and an ANTHROPIC_API_KEY. Detailed steps, schema, and environment variables will be documented in the wiki.
The project is built and live. Ongoing work is focused on:
- A guided, one-command setup so anyone can deploy their own instance
- A richer community layer (discussion, contributor recognition)
- Optional retrieval scaling for very large source libraries
- Streaming responses, without ever weakening the honesty checks
Contributions are welcome see an opportunity, open an issue or a pull request.
Released under the MIT License — free to use, modify, and distribute, including commercially. The one requirement: keep the original copyright and license notice. See LICENSE.
Copyright © 2026 Wael Alebrahim.