A personal AI knowledge curation platform with contradiction intelligence.
Noesis turns scattered tweets, blogs, and videos into structured knowledge you can retrieve quickly and challenge for contradictions.
Entry 1: Tips on writing your first research. Entry 2: From an idea to an ML paper in under.
git clone https://github.com/garg-tejas/noesis.git
cd noesis
pnpm install
cp .env.local.example .env.local
# add Supabase + Gemini values
# run SQL files in scripts/ using Supabase SQL editor (in order)
pnpm devOpen http://localhost:3000.
Saving content is easy. Building reliable understanding is hard.
Read-later workflows usually fail because:
- high-value signal is buried in low-value content
- stored knowledge is hard to retrieve when needed
- contradictory claims accumulate silently
- AI Distillation: Extracts core ideas, context, actionables, and quality scores (0-100) from Twitter threads, blog posts, and YouTube videos
- Smart Retrieval: Server-side search across all entries with filters for source type, tags, and quality threshold
- Contradiction Intelligence: Analyzes semantically related entries (grouped by tags) and surfaces conflicting claims with explanations
- Personal Layer: Favorite entries, add notes, and track contradiction timeline in your knowledge base
| Endpoint | Purpose |
|---|---|
POST /api/distill |
Extract structured knowledge from raw content |
GET /api/entries |
Server-side search, filters, and pagination |
POST /api/contradictions |
Analyze entries for conflicting claims |
GET /api/stats |
Dashboard analytics (quality, sources, tags) |
GET /api/dashboard |
Single response: entries + stats (dashboard bootstrap) |
All endpoints enforce auth + input validation.
- Contradiction detection algorithm: avoids expensive all-pairs comparison by pre-grouping entries with shared tags, then filters out self-references and invalid pairs before sending to Gemini.
- Server-side search with pagination: query params are validated in
GET /api/entries, then filtered by search text, tags, source, and quality. - Rate limiting on AI-heavy routes: fixed-window limits on
/api/distilland/api/contradictionskeyed by user + client IP. - Canonicalized persistence: contradiction pairs are normalized (
item1_id < item2_id), deduped in service logic, and reinforced with SQL uniqueness constraints.
- Framework: Next.js 16 (App Router), React 19, TypeScript
- Data + Security: Supabase Postgres, Supabase Auth, RLS-enabled core tables
- AI: Google GenAI (Gemini)
- UI + Deploy: Tailwind CSS, Radix UI, Vercel
- Contradiction detection still requires human judgment to validate claims
- Quality scores depend on model prompt quality and can drift
- Optimized for personal use (single-user); not designed for team collaboration yet
- No PDF or academic paper ingestion (text/YouTube only)
- The hard part is not just generating summaries; it is preserving retrieval quality as the dataset grows.
- Contradiction detection becomes noisy and expensive without strong pre-filtering before model calls — naive all-pairs comparison on 100 entries would cost 4,950 API calls; tag-based grouping reduces this to ~50.
- AI-backed write paths need idempotency primitives (canonical ordering + dedupe) to keep persistence clean.
- Reliability comes from boundaries: schema validation, timeout/retry handling, and route-level throttling.