TMCandClaude Opus 5.5 4789767103
Release / build (push) Successful in 1m4s
Group stories with the triage model when the server has no embeddings
Chat-only OpenAI-compatible servers (API proxies) answer /embeddings with
404. Every kept article then failed at the grouping step, so nothing was
ever summarized or published.

- Grouping now has two engines: embeddings (as before) and the triage
  model, which gets the new article and a numbered list of the last 48 h
  of stories and picks the same event or none. It works across languages.
- The triage-model engine is used when no embedding model is set, and
  automatically after /embeddings answers 400/404/405/422/501 for the
  configured endpoint and model (remembered until LLM settings change),
  so the article that hit the 404 is grouped in the same step.
- Test connection explains a missing embeddings endpoint instead of a bare
  "HTTP 404"; an empty embedding model is no longer reported as a failure.
- Settings copy: the embedding model is optional.
- Mock LLM: option for no embeddings endpoint, and story-match answers;
  new test reproducing a chat-only server with a chat model in the
  embedding field.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-15 11:50:03 +01:00

RSSFeed

A self-hosted news pipeline. You add RSS/Atom sources; RSSFeed fetches them, has your local LLM read every article, removes what you said you never want to see, groups reports of the same event into one story, and writes a one-sentence Croatian title and a description of at most two sentences for each story. The result is one merged RSS feed (plus a dark-mode web reader), at the volume you choose.

The original articles are never edited or republished: every item links to the article exactly as published.

  • Negatives in plain language ("celebrity gossip", "horoscopes"), plus free keyword rules and "always keep" exceptions. Every rejection is listed with its rule and reason, and can be rescued.
  • Stories, not duplicates: the same event from five outlets becomes one item listing every source, across languages.
  • Importance 1–5 (neutral scale), feed controls (daily limit, editions, minimum importance, per-source and per-category caps, with a 7-day preview), and a daily digest ("Jutarnji pregled").
  • Standard Croatian, checked: a format check enforces the sentence limits and a language guard blocks Serbian/Bosnian forms and untranslated text.

The full product spec is in PRD.md.

Run it

docker login git.mariosem.es          # once; the image is in the private registry
docker compose pull && docker compose up -d

Open http://<host>:34500, then:

  1. Settings → LLM connection: the base URL of any OpenAI-compatible server (LM Studio, Ollama, llama.cpp, vLLM, …), for example http://192.168.1.50:1234/v1. Pick a triage model (small and fast is fine) and a summary model (the best you have for Croatian). An embedding model is optional: a multilingual one such as bge-m3 groups stories fastest, but if your server has no /embeddings endpoint (chat-only proxies), leave it empty and the triage model groups stories instead. RSSFeed also switches to that automatically when /embeddings answers 404. Press Test connection.
  2. Sources: paste a site or feed URL; the feed is found for you. OPML import works too.
  3. Rules: add your negatives. Use Dry run to see what a rule would remove before saving it.
  4. Settings → Feed controls: choose Live or Editions and how many stories a day you want. The preview replays the last 7 days with the new settings.
  5. Settings → Playground: paste any article URL to see the extraction, the verdict and the title/description, and compare two models side by side.

Feeds

URL Content
/feeds/all.xml RSS 2.0: stories admitted by feed controls, plus the daily digest
/feeds/all.json The same as JSON Feed 1.1
/feeds/digest.xml Daily digests only

Query parameters narrow a subscription: ?min_importance=4, ?category=Tehnologija, ?source=<id>, ?limit=50.

Behind Pangolin: RSS reader apps cannot complete a browser login, so give the /feeds/ path its own Pangolin rule. If you open it without a login, set RSSFEED_FEED_TOKEN and the feeds move to /feeds/<token>/all.xml. Set RSSFEED_PUBLIC_URL to the public address so the feeds' self links are right (otherwise they are derived from the proxy's forwarded headers). The app has no login of its own; never publish its port straight to the internet.

Configuration

Everything else lives in the UI (Settings). Environment variables:

Variable Purpose Default
RSSFEED_PORT HTTP port 34500
RSSFEED_DATA_DIR Data directory (SQLite database) /data
TZ Timezone for the digest, editions, daily limits and displayed times Europe/Zagreb
RSSFEED_LLM_BASE_URL Initial LLM endpoint (editable in the UI) none
RSSFEED_LLM_API_KEY Optional key for the LLM endpoint none
RSSFEED_PUBLIC_URL Public base URL through Pangolin, for feed self links from forwarded headers
RSSFEED_FEED_TOKEN Serve feeds only under /feeds/<token>/ none
RSSFEED_ALLOW_PRIVATE_URLS Allow source URLs on private/LAN addresses (blocked against SSRF) off
RSSFEED_LOG_LEVEL debug, info, warn, error info

Logs are JSON lines on stdout. GET /api/health reports version, database, queue depth, LLM state and the last edition/digest times.

Develop

Requires Node 24 (TypeScript runs through Node's native type stripping; no build step for the server).

npm install
npm run dev -- --mock     # server :34500 + web :5173 + mock LLM :11435 → http://localhost:5173
npm run dev               # same, with your own LLM configured in Settings
npm test                  # unit tests + end-to-end pipeline test against the mock LLM
npm run typecheck
npm run build             # web UI → web/dist (served by the server in production)

The mock LLM (npm run mock-llm) answers every request shape with deterministic Croatian output, so the whole pipeline can be exercised without a GPU. Its models are mock-small, mock-large and mock-embed.

Layout: server/ (Express 5, node:sqlite, pipeline), web/ (React 19 + Vite), shared/types.ts (the API contract used by both).

Release

.gitea/workflows/release.yml builds the image on every push to main (pull requests build only). The Dockerfile's build stage runs the typecheck and the full test suite, so a failing test never produces an image. Images are scanned with Trivy and pushed as git.mariosem.es/msemes/rssfeed:latest and :<short-sha>.

The workflow needs a repository secret REGISTRY_TOKEN: a Gitea personal access token with write:package scope.

S
Description
No description provided
Readme
774 KiB
0 Stars 1 Watchers 0 Forks
Languages
TypeScript 92.7%
CSS 5.3%
JavaScript 1.4%
Dockerfile 0.4%
HTML 0.2%