News & daily briefing¶

Jarvis builds a personalised briefing of the day's news, deduplicated, AI-summarised, delivered via chat / email / TTS audio.

What you can do¶

• 📰 Daily briefing AI-generated with the top stories relevant to you
• 🎯 Topic filters (tech, crypto, finance, health, city, hobbies)
• 🔄 Deduplication across different sources covering the same event
• 🌍 Multilingual: read in IT, get the summary in IT even if the source is in EN
• 🎙️ Audio briefing via TTS for car / running consumption
• ✅ Source citations always included

Recommended stack¶

Self-hosted RSS readers¶

Recommended: Miniflux — single binary, lightweight, clean REST API, perfect Jarvis backend.

Commercial news APIs¶

Decentralised networks¶

• Bluesky (AT Protocol) — public feeds without auth
• Mastodon — ActivityPub REST per instance
• Lemmy — self-hostable Reddit alternative

Podcasts¶

• Castopod — self-hosted podcast server
• AntennaPod + OPML sync

Daily briefing architecture¶

flowchart TB
    subgraph Ingestion
        MF[Miniflux RSS] & GDELT[GDELT] & CURR[Currents API] & BSKY[Bluesky] & MAST[Mastodon]
    end
    Ingestion --> DEDUP[Deduplication\nheadline normalisation]
    DEDUP --> FILTER[Topic filter\nembeddings BGE-M3]
    FILTER --> RANK[Ranking\nrecency · trust · interest]
    RANK --> SUMM[Summarisation\nmap-reduce LLM]
    SUMM --> FACT[Fact-check\ncross-source]
    FACT --> RENDER[Briefing render]
    RENDER --> CHAT[Chat]
    RENDER --> EMAIL[Email]
    RENDER --> TTS[Audio Piper TTS]

Technical pipeline¶

1. Parallel pull every N minutes from configured sources
2. Deduplication: title normalisation → word-overlap > 70% = duplicate
3. Topic filter: BGE-M3 embedding (multilingual) + cosine similarity with your interest vectors
4. Ranking: score = recency × source_trust × user_interest
5. Summarisation:
6. extractive (sumy / newspaper3k) → 80% reduction
7. abstractive (Mistral/Qwen via Ollama, or Claude/GPT)
8. map-reduce: summary per article → summary of summaries
9. Optional fact-check: cross-reference between sources, flag "unconfirmed"
10. Render: structured JSON → Jinja2 template → chat/email/TTS

Configuration¶

# Miniflux self-hosted
MINIFLUX_URL=http://miniflux:8080
MINIFLUX_API_TOKEN=...

# News API
CURRENTS_API_KEY=...

# Bluesky (public read, no auth required)
BLUESKY_FEEDS=at://did:plc:.../app.bsky.feed.generator/whats-hot

# Audio briefing TTS
PIPER_VOICE=en_US-amy-medium

# config/jarvis.yaml
news:
  briefing:
    schedule: "0 7 * * *"   # every morning at 07:00
    topics:
      - tech
      - ai
      - crypto
      - markets-eu
    sources:
      - miniflux:starred
      - currents:tech
      - bluesky:whats-hot
      - gdelt:gkg-global
    output:
      - chat
      - email: your@email.com
      - audio: smartwatch
    sections:
      - top_stories: 3
      - trends: true
      - markets: true
      - calendar: true

Usage examples¶

Morning briefing¶

"Hey Jarvis, briefing!"

Jarvis: Good morning. Three top stories:

1. ⚡ Anthropic releases Claude Sonnet 4.7 — 2x latency, same prices
   (sources: Anthropic, TheVerge, HackerNews)

2. 🇪🇺 ECB holds rates, EUR/USD at 1.085
   (sources: Reuters, FT)

3. ₿ Bitcoin stable at 95,400 USD, spot ETF inflows +1.2B

Today's trends: Python agent frameworks growing.
Markets: SPX -0.2%, FTSE 100 +0.5%.
Calendar: internal meeting at 11:00, dentist at 16:30.

Audio or text only?

Follow-up questions¶

"You mentioned agent frameworks, tell me more" "What did HackerNews say about Anthropic?"

Ethical filtering¶

• 🚫 Auto-skip clickbait (titles matching "you won't believe X", "Y something")
• ⚖️ Source balancing (not a single ideological bubble)
• ⚠️ "Opinion" vs "fact" labelling when detectable
• 📊 Trust score per source (Mediabias/Factcheck)

Privacy¶

• ✅ The whole pipeline can run 100% locally (Miniflux + Ollama)
• ❌ Cloud LLM: article content passes through the provider
• 🌐 GDELT, Currents, Bluesky require no user identification

Roadmap¶