Flashcard Generator Powered by GPT

Overview

Generates study flashcards from arbitrary text input using an LLM backend, with RAG to ground cards in user-supplied material.

Technical Challenges

RAG Pipeline

User documents are chunked with a sliding window (with overlap to preserve sentence context), embedded via OpenAI embeddings, and stored in a vector index. At generation time, the top-k chunks are retrieved and injected into the prompt as grounding context, reducing hallucination rates for domain-specific content.

Structured Output & Prompt Engineering

GPT doesn't reliably return well-formed flashcard arrays without explicit guidance. The prompt enforces a strict JSON schema (question/answer pairs with optional hints), and responses are validated with Zod before being persisted malformed responses trigger a retry with an error-correction prompt.

Authentication & Data Isolation

Each user's card decks are scoped by session, with server-side auth checks on every API route to prevent cross-user data access.

Latency

Generation for large documents can exceed typical request timeouts. Addressed by streaming the LLM response and flushing cards to the client incrementally as each question/answer pair is parsed.