Overview

AI hallucinates fake references. ChatGPT and other LLMs generate up to 40% fabricated citations. The Source Taster catches them — validated in a Master's thesis at the University of Siegen.

The system helps students and researchers verify bibliographic references in seconds. It extracts references from text or PDFs, searches them across 5 academic databases, and scores each match transparently — field by field.

Key Metrics

• 93% APA exact match rate
• 100% hallucination detection on synthetic fakes
• <3 seconds average verification time per reference
• 97.2% F1-score on curated test sets (n=425 survey + automated benchmarks)

How It Works

1. Import — Paste text, upload a PDF, or right-click a bibliography on any page.
2. Extract — AI parses unstructured references into structured CSL-JSON (title, authors, DOI, year, etc.).
3. Search — Each reference is queried against OpenAlex, Crossref, Semantic Scholar, Europe PMC, and arXiv.
4. Match — Deterministic scoring compares extracted data with database candidates. Results show a colour-coded per-field breakdown.
5. Verify — Green means verified. Red means suspect. You know instantly which sources are trustworthy.

Architecture

The system consists of three main building blocks:

• Browser extension (apps/extension) — Vue 3 application with Vuetify that imports references (text, PDF, context menu), performs AI-assisted extraction, and visualises verified matches.
• API (apps/api) — Hono server on Node.js 20 that orchestrates AI extraction, AnyStyle parsing, database searches, and deterministic matching.
• Shared types (packages/types) — Zod schemas and TypeScript definitions for CSL-JSON, API contracts, matching configuration, and UI defaults. Both the extension and API consume them.

Extension Highlights

• Import via context menu ("Check bibliography"), PDF upload, or manual text paste.
• AI extraction with configurable CSL fields taken from shared defaults.
• AnyStyle-powered token labelling and manual correction before converting to CSL.
• Database search with user-defined priorities and optional early termination for high scores.
• Field-level match visualisation with colour-coded scores directly inside the UI.

API Highlights

• /v1/extract — AI extraction with response_format=json_schema and automatic fallback to json_object.
• /v1/search/:database — provider-specific search with DOI/identifier shortcuts and query heuristics.
• /v1/match — deterministic matching with normalisation and similarity heuristics.
• /v1/anystyle/* — proxy to a Ruby AnyStyle server for tokenisation and CSL conversion.
• /v1/user/ai-secrets — store, inspect, and delete user keys (scoped by X-Client-Id).

A full flow from extraction to matching is illustrated in Architecture.