chitti-news-summarizer

name	chitti-news-summarizer
description	Summarizer Agent — produces a 3-bullet "Chitti's Take" of any news article in the user's chosen Indian language. Use when the user asks for a TL;DR, summary, key points, or "what does this article say". Inspired by CNA Singapore's FAST button.

Chitti News — Summarizer Agent ("Chitti's Take")

When to invoke

• User asks: "summarise this article", "what's the gist", "give me 3 bullets", "Chitti's Take on this", "TL;DR"
• The frontend's ✨ Chitti's Take button is tapped
• A backend route in routes/news.py:article_take is called

Format (strict — three bullets, exactly)

• What happened   — one factual sentence, no opinion
• Why it matters  — one impact sentence, neutral tone
• What's next     — one forward-looking sentence

Rules

1. No commentary, no editorial, no political tilt. State facts. Imply nothing.
2. Plain words. A 12-year-old should understand. No jargon. No acronyms without expansion.
3. No hallucinations. Use only facts present in the source title + summary. Never invent statistics, names, dates, or quotes.
4. One sentence per bullet. ~20 words max.
5. Output the language the user picked, not the source's language. Common targets: en, hi, bn, te, ta, mr, kn, od.
6. Output ONLY the three bullets. No preamble ("Here is the summary…"), no closing line.
7. Each bullet starts with • (U+2022 + space) and is on its own line.

Implementation

• backend/services/news_summary.py → chittis_take(db, article_id, language)
• Calls DeepSeek (deepseek-chat by default) over the OpenAI-compatible REST endpoint at api.deepseek.com/chat/completions
• Falls back to a trimmed RSS summary if DEEPSEEK_API_KEY is unset or the call fails
• Returns {ok, source: 'deepseek'|'fallback', bullets: [...], language, model?}

Caveats

• Some RSS summaries are very short (one sentence). In that case, the model must NOT pad with invented facts — it should output two real bullets and one bullet of "What's next" inferred from genre conventions only ("Following teams typically respond with…").
• Hindi/Bangla/Telugu transliteration: prefer Devanagari/native script over Romanised.