PulseSphere AI — Local LLM Edition

What changed

Area	Before	After
Summarization	OpenRouter API (Mistral 7B cloud)	Local BART (facebook/bart-large-cnn) — no API key
Embeddings	Simple hash-based numpy vectors	sentence-transformers all-MiniLM-L6-v2
Vector DB	In-memory numpy array	ChromaDB (persistent on disk at ./chroma_db)
New endpoint	—	GET /api/semantic-search?q=...
New endpoint	—	GET /api/vector-db/stats
Frontend	Card grid only	Card grid + Live News Feed section
Summarize UX	—	"Summarize" button on every card + feed item → modal

---

Backend setup

cd backend

# 1. Create .env
cp .env.example .env          # add your GNEWS_API_KEY

# 2. Install deps
pip install -r requirements.txt

# CPU-only PyTorch (saves ~2 GB):
pip install torch --index-url https://download.pytorch.org/whl/cpu

# 3. Run
uvicorn server:app --reload --port 8000

First summarize request: BART (~1.6 GB) downloads automatically to ~/.cache/huggingface/. Subsequent requests are instant.
If BART is unavailable (low RAM), the service falls back to extractive summarization automatically.

---

Frontend setup

cd frontend
npm install
REACT_APP_BACKEND_URL=http://localhost:8000 npm start

---

New UI sections

Live News Feed

• Scrollable list of latest headlines at the bottom of the page
• Shows thumbnail, category badge, source, time-ago
• Hover any article → Summarize button appears
• Click Summarize → AI summary modal (runs local BART)
• Auto-refreshes every 5 minutes
• Badge shows how many articles are indexed in ChromaDB

Summarize on Cards

• Every card in the grid also has a Summarize button
• Same modal experience — title, summary, link to full article

---

Architecture

Browser
  │
  ├─ GET /api/news          → news_service.py (GNews API)
  │                           ↳ background: ChromaDB.upsert(articles)
  │
  ├─ GET /api/search        → news_service.py (GNews search)
  │                           ↳ background: ChromaDB.upsert(results)
  │
  ├─ POST /api/summarize    → ai_service.summarize_text()
  │                           ↳ BART (local) → extractive fallback
  │
  ├─ GET /api/semantic-search?q=… → ChromaDB.query()
  │
  └─ GET /api/vector-db/stats    → ChromaDB.count()

Local LLM Stack