🍊 HN Doom-Scroll

A self-hosted Hacker News reader that summarizes every article with AI so you can skim the front page in minutes instead of hours.

Most Hacker News clients just restyle the same orange list. HN Doom-Scroll actually reads each article for you — extracting full text via headless browser, generating a 2-sentence summary, and presenting everything as an infinite-scroll feed you can fly through. Mark stories read, save them, or hide them — and they never come back. It even learns what you skip and quietly demotes similar stories over time.

Search everything you've ever seen by keyword or semantic meaning. Runs on AWS Bedrock (fast, no GPU needed) or fully offline with local Ollama models. No accounts, no tracking, no ads — just your own private Hacker News experience.

Use cases: daily HN catchup, research triage, building a personal knowledge base from Hacker News articles, staying current on tech without the time sink.

Why build another Hacker News reader?

Existing HN clients — Hacker News apps, RSS readers, Algolia-based tools — show you the same title + score + comment count. You still have to click through to every article to decide if it's worth your time.

HN Doom-Scroll solves this differently:

1. AI reads the article for you. Full text extraction (even JS-heavy pages) plus a concise summary means you know what's in the article before clicking.
2. Personal relevance learning. After a few hides, recurring topics you don't care about get pushed to the bottom automatically.
3. Semantic search across your history. Find that article about "zero-copy networking" you saw last week — even if those exact words weren't in the title.
4. No vendor lock-in. Switch between cloud AI (AWS Bedrock) and local models (Ollama) with one toggle. Your data stays in a local SQLite file.

How it compares to other Hacker News tools

Tool	Summaries	Learns preferences	Semantic search	Self-hosted	Offline capable
HN Doom-Scroll	✅ AI-generated	✅ Auto-learns	✅ Embeddings	✅	✅ (Ollama)
Official HN	❌	❌	❌	❌	❌
Hacker News apps (iOS/Android)	❌	❌	❌	❌	❌
HN Algolia search	❌	❌	❌ keyword only	❌	❌
RSS readers	❌	❌	❌	varies	varies
Daily.dev / similar	some	basic	❌	❌	❌

---

Screenshots

Light	Dark

Settings — provider toggle, model selection, themes, and filters:

AI Provider settings — switch between Bedrock and Ollama with model dropdowns:

Semantic search ranks results by meaning, with a match score on each:

Regenerate screenshots: python scripts/capture_screenshots.py and python scripts/capture_extra.py (requires a running instance and Playwright).

Stack

Layer	Technology
Backend	Python + FastAPI
Storage	SQLite (hn.db, auto-created)
Summaries	AWS Bedrock (google.gemma-3-4b-it, default) or local Ollama (llama3.2:3b)
Embeddings	AWS Bedrock (amazon.titan-embed-text-v2:0, default) or local Ollama (nomic-embed-text)
Article extraction	trafilatura + Playwright headless Chromium fallback
Frontend	Vanilla HTML/CSS/JS, infinite scroll, no build step

Quick start

Prerequisites

Option A — AWS Bedrock (default, fast)

1. Configure AWS credentials with Bedrock access (aws configure or env vars).
2. Double-click run.bat (or run from a terminal).
3. Browser opens to http://localhost:8000.

No model downloads needed. Both summaries and semantic search work immediately via Bedrock (Gemma 3 4B for summaries, Titan Embed v2 for search).

Option B — Local Ollama (fully offline, no cloud dependency)

1. Install and start Ollama.
2. Pull models:

   ollama pull llama3.2:3b
   ollama pull nomic-embed-text   # optional — enables semantic search
   

3. Set the env var before launching:

   set HN_PROVIDER=ollama
   run.bat
   

Summaries are slower (CPU-bound) but everything stays on your machine.

First run

run.bat creates the virtualenv, installs dependencies (including boto3 for Bedrock), and downloads Playwright Chromium (~150 MB, one-time). Playwright is optional — skip it and the app falls back to direct fetch + HN discussion summaries for JS-heavy pages.

How it works

• Fetches HN top stories on startup and stores them in SQLite.
• On-demand summaries. As each card scrolls near the viewport, the app calls the configured provider to summarize that article. The header bar shows a running count of pending summaries. Cards display "⏳ Summarizing…" until their summary arrives, then update in place.
• Article extraction is layered. For each story: (1) direct fetch with browser headers, (2) headless Chromium render for JS-heavy pages, (3) fallback that summarizes the HN discussion. Badges show the source — 💬 From HN discussion, 🌐 Rendered page — or why it couldn't be read — 🔒 Paywalled, 📄 PDF, 🎥 Video.
• Crash recovery. If the server is killed mid-summarization, stuck pending stories are automatically reset on the next startup.
• ✓ Read / ★ Save / ✕ Not interested — each moves a story out of the feed. Stories never reappear once acted on.
• Learns from what you skip. After 10+ hidden stories, recurring patterns (words, domains) down-rank similar new stories to the bottom of the feed, dimmed with a reason. Nothing is hidden outright.
• Bulk-hide dimmed stories. A "Hide all dimmed stories" banner appears between your normal stories and the down-ranked ones. One click hides them all at once.
• Search across everything (feed, read, saved, hidden) by keyword or semantic meaning. Toggle kw/ai in the search box. Semantic search uses the configured provider's embedding model (Titan Embed v2 on Bedrock, nomic-embed-text on Ollama). Query embeddings are cached for 10 seconds to avoid redundant API calls during typing. Embedding backfill runs in the background after summaries complete, keeping search responsive.
• Auto-refresh pulls the latest front page on a timer. The refresh button doubles as a countdown.

Settings

Click ⚙ in the header. Everything persists between sessions.

• AI Provider: toggle between ☁️ Bedrock (cloud) and 💻 Ollama (local). Switching takes effect immediately for new summaries and searches.
• Bedrock Models: (shown when Bedrock is active) select from curated cost-effective models for summarization (Claude 3 Haiku, Gemma 3 4B, Llama 3 8B, Nova Micro) and embeddings (Titan Embed v2, Cohere Embed v3/v4).
• Ollama Models: (shown when Ollama is active) pick from models installed locally. Pull more with ollama pull <name>.
• Themes: Light, Dark, or System (follows OS preference).
• Auto-refresh: 5–60 min interval, countdown in the refresh button.
• Feed size: 25–200 top stories per refresh.
• Keyword filters: hide stories containing specific words.
• Re-embed all: regenerate all stored embeddings after switching models.

Provider configuration

Env var	Values	Default
HN_PROVIDER	bedrock or ollama	bedrock
BEDROCK_REGION	AWS region	us-east-1
BEDROCK_REASON_MODEL	Bedrock model ID for summaries	google.gemma-3-4b-it
BEDROCK_EMBED_MODEL	Bedrock model ID for embeddings	amazon.titan-embed-text-v2:0

In Ollama mode, models are selected in-app via Settings → Models.

Files

File	Purpose
app.py	FastAPI app, routes, on-demand summary + embedding generation
hn.py	Hacker News API client
summarizer.py	Article extraction, summarization, and embeddings (Bedrock + Ollama)
db.py	SQLite schema, queries, search, dislike-learning
static/	Frontend (index.html, style.css, app.js)
scripts/capture_screenshots.py	Screenshot generator for README
scripts/capture_extra.py	Additional screenshots (semantic search, provider settings)
docs/	README screenshots
run.bat	One-click launcher
push-to-github.bat	Publishes to standalone GitHub repo
requirements.txt	Python dependencies
CHANGELOG.md	Version history
LICENSE	MIT

Notes

• The status indicator in the header shows green when the provider is reachable, red when it isn't.
• Bedrock mode runs summaries concurrently (fast). Ollama mode serializes them (one at a time, to avoid overloading the local model).
• In Ollama mode, semantic search requires the nomic-embed-text model. If it isn't installed, search falls back to keyword automatically.
• Ask HN / Show HN text posts are summarized from the post + discussion directly.