Overview

PDF Chat is an early prototype I built to learn retrieval-augmented generation (RAG). You upload a PDF, the app indexes it, and you can then ask questions in a chat window and get answers grounded in the document's own text rather than the model's general knowledge.

My role

Solo project. I wrote the Next.js client and two iterations of the Express backend, the later one with documented upload and chat endpoints.

What it does

• Upload: accepts PDF files only, up to 10 MB, and deletes the temporary file once it has been indexed.
• Index: extracts the text, splits it into roughly 1,000-character chunks with 200 characters of overlap, and embeds each chunk with OpenAI embeddings.
• Store: saves the vectors in a FAISS index on disk, so the document survives a server restart.
• Answer: for each question it retrieves the four most relevant chunks and passes them to an OpenAI chat model at low temperature through a LangChain retrieval QA chain.
• Chat UI: a single page with upload status, a message thread and readable error messages.

How it's built

The front end is a Next.js app whose API routes act as a thin proxy: the upload route parses the multipart form and forwards the file, and the chat route forwards the question. The backend is Express with Helmet, Morgan and Multer, using LangChain's PDF loader, recursive text splitter, OpenAI embeddings, FAISS vector store and retrieval QA chain. Upload and chat are separate routes, and the vector store is loaded lazily from disk on the first question.

Highlights

• Persistent local vector index: FAISS runs in-process with no external database, and the saved index is reloaded on demand.
• Clean layering: routes, services and a small vector-store module are kept separate, so the retrieval logic can change without touching the HTTP layer.