How VeriLens Works
Detailed breakdown of the IndoBERT model, the 3-stage RAG pipeline, and the credit refund guarantee.
Architectural Philosophy
Misinformation verification requires balancing sub-second throughput against rigorous investigative depth. VeriLens addresses this through a dual-engine architecture:
Tier 1: IndoBERT Fast Classification
Trained on Indonesian social media hoaxes, news articles, and verified datasets, our IndoBERT model evaluates syntactic structures, clickbait rhetoric, sensationalism, and linguistic anomalies directly on-premise without external search queries. Ideal for high-volume message filtering, chat bots, and instant mobile previews (<100ms latency).
Tier 2: Stage 1 — Claim Extraction & Normalization
When calling `/api/v1/predict/deep`, Stage 1 extracts atomic assertions, key entities (politicians, government bodies, brand names), and temporal anchors from colloquial or forwarded text, generating an unambiguous normalized claim.
Tier 2: Stage 2 — Dual-Strategy Evidence Retrieval
Stage 2 generates dual verification queries: 1. **Verification Queries**: Targets existing debunking coverage in TurnBackHoax, Mafindo, and Kominfo databases for known viral hoaxes. 2. **Source-Lookup Queries**: Extracts distinctive non-suspicious entities to locate the authentic original news report if an authentic article was maliciously altered. Search results are deduplicated, filtered for freshness, and bounded to reliable domains.
Tier 2: Stage 3 — Grounded Verdict Evaluation
Using Google Gemini, Stage 3 cross-references the retrieved evidence snippets against the normalized claim. The model produces: • A definitive label: `fakta`, `hoaks`, or `tidak_dapat_diverifikasi`. • A calibrated confidence score (`0.0 - 1.0`). • Nuanced Indonesian reasoning explaining why the claim holds or fails. • Direct clickable URLs of verified sources cited in the reasoning.
Credit Deductions & Auto-Refund Guarantee
Deep analysis consumes 1 credit per execution. If network disruption, web crawler timeouts, or upstream inference anomalies occur, the deduction is automatically rolled back and the credit is immediately refunded to the user's account.