Dispatch

Dispatch is intended to be a centralized aggregation of tech news.

Start-up news, major AI labs, and other general tech information all condensed into one email. Additionally, Dispatch has a RAG system to answer questions over the same index w/ citations to answer any questions about articles or other general questions regarding the news listed.

Repo layout

design.md                     landing-page spec (copy, tokens, component tree)
frontend/                     Vite + React + TypeScript + Tailwind — landing, auth, chat, profile
backend/                      FastAPI + SQLAlchemy + pgvector RAG service (see backend/README.md)
.github/workflows/ingest.yml  nightly ingestion cron

Stack

• Frontend: React, Vite, TypeScript, Tailwind CSS, lucide-react, React Router
• Backend: Python 3.13, FastAPI, async SQLAlchemy, Postgres + pgvector, Alembic
• Auth: Supabase Auth (email/password + Google OAuth), JWT-verified on the API
• Models: OpenAI GPT-4.1 Nano (generation, classification, summaries) · text-embedding-3-small @ 1024-dim (embeddings)

How it works

Ingestion → index

Nightly, a scheduled GitHub Actions workflow (.github/workflows/ingest.yml) runs the pipeline across the live sources:

1. Fetch — TechCrunch, Hacker News, arXiv, and AI-lab blogs (Anthropic, OpenAI; DeepMind planned)
2. Deduplicate — URL-fingerprint check drops redundant articles before any LLM cost
3. Summarize — concise per-article summaries
4. Classify — topic + signal score, enabling fine-grained filtering downstream
5. Embed & store — chunk embeddings written into pgvector (HNSW, cosine)

RAG chat

Hybrid retrieval with streamed, cited answers:

• Vector search — cosine similarity over HNSW-indexed chunk embeddings with a recency-decay penalty
• Keyword search — Postgres full-text search via websearch_to_tsquery over a GIN-indexed tsvector column (title weight A / summary weight B); supports natural-language queries, quoted phrases, and negation
• Fusion — Reciprocal Rank Fusion (RRF, k=60) over both ranked lists
• Generation — streamed (SSE) answers with inline [n] citations mapped back to source items

Accounts & persistence

• Auth — Supabase sign-in (email/password + Google); the API verifies the Supabase JWT on every chat request
• Chat history — conversations are saved per user: deep-linkable (/chat/:id), restored on refresh, synced across devices, with model-generated titles (via same GPT model)
• Profile — editable display name and topic-preference selection

Quick start

The frontend and backend run separately.

Backend (details in backend/README.md):

cd backend
uv sync
alembic upgrade head
uvicorn app.main:app --reload

Frontend:

cd frontend
npm install
npm run dev

Deployment TBD (To Be Deployed): planned via Vercel (frontend) + Render (backend).