PDFSnap is now public!
Chat with your documents in seconds.
PDFSnap lets you have real conversations with any PDF. Upload your file and start asking questions instantly.
Get started
Production-grade AI, built from scratch
by Mohamed Aziz Nacib
Every layer was individually integrated — from vector ingestion to billing webhooks — with production concerns like multi-tenancy, streaming, and edge security in mind.
RAG Pipeline — how a question becomes an answer
Vector Embeddings
Every PDF is split into semantic chunks and passed through OpenAI's embedding model, producing 1 536-dimensional vectors that capture meaning — not just keywords — for precise retrieval.
Pinecone Vector DB
Vectors are upserted into Pinecone with per-document namespaces for clean multi-tenant isolation. Cosine similarity search retrieves the top-k relevant chunks in single-digit milliseconds.
OpenAI GPT — RAG
Retrieved chunks are injected into a carefully engineered system prompt so the model answers from the document itself, not from hallucination. Responses stream token-by-token for instant feedback.
Stripe Billing
Full Checkout and Customer Portal integration with signed webhook handling for subscription lifecycle events. Plan limits are enforced server-side at both the upload and query layers.
Serverless Ingestion
Files land on UploadThing's CDN via presigned URLs. A serverless callback then triggers chunking, embedding, and Pinecone indexing asynchronously — the user never waits for the pipeline.
Edge-enforced Auth
HS256-signed JWTs live in httpOnly cookies. Route protection runs in Next.js middleware at the Edge before any server component executes — zero auth logic leaks to the client.
1 536
Embedding dimensions
k-NN
Vector similarity search
Edge
Auth middleware layer
100%
Custom-built — no boilerplate
