Olds

A news reader that show you non-obvious connections between stories across topics.

🌐 readolds.fyi

---

---

Olds finds connections between news stories that don't belong to the same category. When you read an article about Iranian trade sanctions, the sidebar shows you a related crime story in Jakarta involving the same political actors — and explains why they're connected in plain English. The feed learns from how you read, not what you explicitly rate.

How it works

Articles arrive from NewsAPI and The Guardian every 30 minutes. A Python service extracts named entities (people, places, organizations) and generates a 384-dimensional sentence-transformer embedding for each article. The Go backend stores both and builds an in-memory directed graph where edge weights combine cosine similarity of embeddings with named entity overlap, so two articles are connected both semantically and concretely.

When you open an article, Go traverses the graph in real time, filters for cross-topic neighbors above a similarity threshold, and pushes them to the browser over a WebSocket. An LLM then generates a 1–2 sentence natural language explanation for each connection, streaming results into the sidebar as they resolve. The feed ranking is shaped by implicit signals — dwell time, scroll depth, re-opens — with no explicit ratings.

flowchart LR
    subgraph Ingestion
        A[NewsAPI / Guardian API] --> B[ML Service\nPython · FastAPI]
        B -- entities + embedding --> C
    end

    subgraph Backend
        C[Go · Gin\nArticle Graph]
        C -- graph traversal --> D[WebSocket Server]
        C -- entity overlap + similarity --> E[LLM API\ngpt-5-nano]
    end

    subgraph Frontend
        F[React · Tailwind\nFeed + Sidebar]
    end

    subgraph Data
        G[(Supabase Postgres\n+ pgvector)]
        H[Supabase Auth]
    end

    D -- connections + explanations --> F
    C <--> G
    H <--> F
    H <--> C

Loading

Metrics

Metric	Value
Articles indexed	11,464
Graph edges	953,251 unique edges
Cross-topic connection rate	59.0%
Enrichment success rate	83.5%
Graph traversal latency	p50 382 ms / p95 819 ms / p99 819 ms
WebSocket push latency	p50 0 ms / p95 0 ms / p99 0 ms
ML inference per article	p50 699 ms / p95 2,795 ms / p99 5,606 ms

Runtime metrics are exposed through GET /metrics, including article count, graph edge count, cross-topic ratio, latency percentiles, ingestion counts, and LLM explanation cache size. Detailed historical snapshots remain behind authenticated /stats endpoints.

Local development

git clone https://github.com/parthkotwal/olds.git
cd olds
cp .env.example .env
# Fill in: SUPABASE_DATABASE_URL, SUPABASE_JWT_SECRET, SUPABASE_ANON_KEY,
#          SUPABASE_URL, NEWSAPI_KEY, GUARDIAN_API_KEY, LLM_API_KEY
docker compose up

Visit localhost:3000. The backend runs on :8080, ML service on :8001.

On first start the backend triggers an ingestion run. It takes 2–3 minutes to fetch articles, run them through the ML service, and build the graph. The feed populates automatically; no manual step needed.

---

Courtesy of Parth Kotwal