Truthify extracts claims from text, assigns an estimated reliability level, and helps readers locate each passage in the original content.
Overview
The problemA series of claims can be hard to examine without losing track of its context, especially in long texts or video transcripts.
The solutionTruthify splits content into chunks, submits them to an OpenAI model, and displays claim cards, scores out of ten, and highlights in the original text.
Paste text or enter a YouTube URL; manual text entry remains available if transcript retrieval fails. Two input modes in a single analysis interface.
1 / 4: Enter content. Previous step / Next step
The React client calls a Next.js SSE route, which orchestrates OpenAI analysis and stores results and cache in PostgreSQL through Prisma.
React
Input, progress and results displayNext.js
Streaming, cache and persistenceTypeScript
Chunking, prompts and score normalizationOpenAI
Claim extraction and assessmentA single request would be sensitive to input length and processing time.
Overlapping chunks with explicit limits on text length and the number of chunks.
Analyzing multiple chunks yields results at different times.
A Server-Sent Events route and React hook expose progress and newly extracted claims.
An isolated assessment lacks context if the user cannot find the original passage.
Excerpt position matching and interactive highlighting connected to result cards.
Transcript retrieval depends on external sources and can fail depending on the video.
A cascade of retrieval strategies, transcript caching, and manual text entry when retrieval fails.
Main technical stack
Input workflow, chunk-level progress, overall score and assessment cards linked to the text.
Analyses stored with Prisma in PostgreSQL, with caching based on a hash of normalized text.
Magic-link authentication, account area, Stripe checkout sessions and signed webhook handling.