| name | chitti-news-factcheck |
| description | Fact Checker Agent — cross-references a news article against ≥2 other trusted RSS sources in our DB and returns a verdict (verified / partial / disputed / unverified) with rationale. Use when the user asks "is this true", "fact check this", "how reliable", "any other sources covering this". |
Chitti News — Fact Checker Agent
When to invoke
- User asks: "is this true", "fact check", "any other sources", "how reliable", "trustworthy"
- The frontend's 🛡️ Fact Check button is tapped
- A backend route in
routes/news.py:article_factcheck is called
Verdicts (4 tiers)
| Verdict | Symbol | Color | Meaning |
|---|
verified | ✅ | green | ≥3 distinct trusted sources agree on the headline + key facts |
partial | 🟡 | amber | 2 sources cover it; broad facts match, details differ |
disputed | ⚠️ | red | 1 other source found, headline diverges OR contradicts |
unverified | ❔ | muted | No cross-source corroboration yet (single-source story — could be hyperlocal or just-breaking) |
Algorithm (v1, no LLM required for matching)
- Normalize the article title: lowercase, collapse whitespace.
- Window to the last 48 hours of articles in the same language.
- Score with
rapidfuzz.fuzz.token_set_ratio().
- Cluster at score ≥ 70.
- Count distinct
source_slug values among matches (excluding the original).
- Cache result in
news.fact_checks for 6 hours; re-running returns the cache.
The rationale (one-line EN + one-line HI explanation of the verdict)
is template-rendered in v1 and stored alongside the verdict. A v2
upgrade to DeepSeek-generated rationale is planned for ambiguous
match scores only — see TODO.md and
PROMPTS.md §2.
Trust assumption
Every source we ingest from data/sources.json is treated as "trusted"
(public RSS feeds from established Indian news outlets). The fact-check
verdict is about agreement among sources, NOT about whether any one
source is correct — Chitti News does not editorialise.
Implementation
backend/services/news_factcheck.py → factcheck(db, article_id, force=False)- DB cache key:
(article_id) unique
- Returns
{ok, verdict, symbol, color, confidence, matched_sources, rationale_en, rationale_hi, checked_at}
Caveats
- Title-similarity matching can yield false negatives when two outlets
use very different wordings for the same event (e.g. "RBI cuts repo
rate" vs "Central bank lowers benchmark"). v2 plans: switch to entity
- verb extraction via DeepSeek.
- "verified" is not a truth claim — it's a "many outlets are saying
this" claim. Always render with the disclaimer "verify on the original
source link before sharing".
