Skip to content

sembr

A self-hosted intent radar — Reverse-RAG matching across RSS, NewsAPI, and Twitter.

sembr is a self-hosted intent radar. You describe what you care about once — "monitor Fed policy impact on emerging-market currencies" — and it continuously scans RSS feeds, news APIs, and social streams, matches articles to your intent via semantic vectors, and delivers LLM-analyzed digests from whatever angle you configure. No keyword lists, no hand-tuned filters. Telegram / Discord / Slack channels and a local LLM backend are scaffolded by the plugin seams and on the post-1.0 roadmap.

How it works

Traditional RAG: user query → search documents → answer.

Reverse RAG (sembr): pre-stored intent vector → continuously scan incoming articles → push semantic matches.

You describe what you care about once. sembr does the watching.

RSS Feeds
    │
    ▼
BGE-M3 Embeddings (SiliconFlow API)
    │
    ▼
Qdrant Vector Store
    │  ◄── Intent vectors (stored at creation time)
    ▼
Semantic Matcher (Qdrant query_points per intent)
    │
    ▼
LLM Summary (OpenAI-compatible chat completions)
    │
    ▼
Email Digest (SMTP)

Each intent picks its own schedule — cron-mode (hourly / daily / weekly preset with a per-intent timezone) or event-mode (fire after N matching articles, or after T seconds since the first buffered match).

Status

1.0 — first stable release.

  • RSS ingestion + BGE-M3 embeddings via SiliconFlow API + Qdrant
  • Intent CRUD via REST API, cron-mode and event-mode schedules
  • LLM-generated summaries via any OpenAI-compatible chat-completions endpoint
  • Email digest delivery via SMTP, rendered in each intent's own timezone with matcher-score badges per source
  • Read-only monitoring dashboard with live log SSE
  • Runtime settings editor — write the host .env and recreate the affected containers in place

License

Apache-2.0 · Built by Peakstone Labs