VeriTrust

AI-powered medical misinformation detection. Users submit a piece of health-related content — pasted text, a URL, or a document (PDF/TXT/MD) — and VeriTrust returns an evidence-grounded verdict: Verdadero (true), Falso (false), Parcialmente falso (partly false), or Dudoso (uncertain), with a confidence score, a per-claim breakdown, the biomedical sources consulted, and a medical report explaining the verdict in plain Spanish.

Health misinformation spreads faster than experts can debunk it. VeriTrust automates the first pass of that debunking: it extracts the verifiable medical claims from a text, cross-checks each one against real biomedical literature (Europe PMC, PubMed, openFDA, AEMPS CIMA), derives a verdict from the stance of that literature, and writes up the result — with the honesty to say "Dudoso" and lower its confidence when the literature is thin, contradictory, or silent.

How it works

Analysis is asynchronous: the pipeline chains several LLM calls and literature lookups, so it runs in a background worker instead of the HTTP request. POST /analysis returns a pending id immediately; an arq worker picks the job off Redis, runs the LangGraph pipeline, and the frontend polls GET /analysis/{id} every 2 s (showing the live pipeline stage) until the verdict is ready.

flowchart TB
    UI["Next.js frontend<br/>(Clerk auth)"]

    subgraph WEB["Web process — FastAPI"]
        API["POST /analysis → insert 'pending' row, enqueue job<br/>GET /analysis/{id} → status + result"]
    end

    PG[("PostgreSQL<br/>analyses, claims, sources")]
    REDIS[("Redis<br/>arq job queue")]

    subgraph WORKER["Worker process — arq (python -m app.worker)"]
        PREP["Prepare input<br/>URL / file text extraction"]
        subgraph GRAPH["LangGraph pipeline"]
            EX["1 · Extractor<br/>claims + search queries + drug terms"]
            TR["2 · Translator<br/>clinical English, one batched call"]
            IN["3 · Investigator<br/>evidence retrieval + relevance judge"]
            HE["4 · Health Expert<br/>evidence verdict + medical report"]
            EX --> TR --> IN --> HE
        end
        PREP --> EX
    end

    LLM["LLM provider<br/>Mistral API (compose) · Ollama (local default)"]
    LIT["Europe PMC · PubMed · openFDA · AEMPS CIMA"]

    UI -- "JWT" --> API
    API --> PG
    API -- "enqueue" --> REDIS
    REDIS --> PREP
    EX -.-> LLM
    TR -.-> LLM
    IN -.-> LLM
    HE -.-> LLM
    IN --> LIT
    HE -- "verdict · claims · sources · report" --> PG
    UI -- "poll every 2s" --> API

The four agents

Agent	Model	What it does
Extractor	llama3	Extracts the verifiable medical claims from the input text, plus an English boolean search query and the drug name (if any) for each claim. Structured output.
Translator	translategemma	Translates all claims to clinical English in a single batched call — the literature sources work in English.
Investigator	ministral-3:14b¹	Queries Europe PMC, PubMed, and openFDA for every claim in parallel — plus AEMPS CIMA when the claim names a drug. An LLM judge filters the hits for relevance and tags each source's stance (supports / contradicts). Computes evidence coverage: the share of claims with relevant literature.
Health Expert	llama3.2	Derives each claim's fake probability from the stance of the retrieved literature (Laplace-smoothed, so thin evidence never reaches certainty), and bands each claim into true, false or uncertain — a band around the decision threshold maps to uncertain, and a claim no source speaks to is uncertain by construction, never false. The overall verdict follows from the claims' verdicts: true and false claims together make it partly false. Then it writes the medical report with llama3.2, grounded in the retrieved sources.

¹ The ministral-3:14b judge needs ~11 GB loaded, ideally on a 16 GB GPU.

Three guardrails temper the raw verdict (app/core/verdict.py):

• Evidence attenuation — confidence is scaled down when little of the input is covered by actual literature, so the system never sounds sure about claims nobody has studied.
• Silence is not falsehood — a claim no retrieved source speaks to comes out Dudoso by construction, never Falso, and Laplace smoothing keeps thin evidence away from certainty.
• A truth never offsets a falsehood — a hoax wrapped in true statements comes out Parcialmente falso, not Verdadero (ADR-0002).

Product surface

• Analysis of pasted text, URLs (SSRF-hardened fetching), and uploaded PDF/TXT/MD files (≤ 10 MB), with per-stage progress while pending and retry for failed runs.
• Per-claim breakdown — each extracted claim gets its own verdict and confidence, alongside the sources that support or contradict it.
• History with search, pagination, and CSV export (formula-injection safe).
• Dashboard with aggregate metrics over your analyses.
• Public share links — a read-only /r/{token} page per analysis, revocable.
• Auth & limits — Clerk JWT on every API route, per-user rate limiting on submission.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗