Chat with PDFs using KNN, not a vector database

What it does pdfGPT is a web application that chunks uploaded PDFs, generates embeddings for each chunk using a Deep Averaging Network Encoder, and retrieves the top matches with K-Nearest Neighbour search to feed into an OpenAI prompt. The goal is to let you ask questions about a document and receive answers that cite page numbers in square brackets. It has been around since 2021, making it an early open-source attempt at what later became known as RAG.

The interesting bit Instead of following the modern playbook—vector database, heavy indexing, dense retrieval servers—the author deliberately keeps the stack minimal: embeddings are computed, saved to disk, and queried with KNN. The author claims this bare-bones approach is why accuracy has held up even as the ecosystem grew fancier.

Key highlights

• Uses the Universal Sentence Encoder family (Deep Averaging Network) rather than OpenAI embeddings for chunk similarity.
• Retrieves context with KNN and injects the top chunks into the prompt, citing source page numbers.
• Supports GPT-4, 16K, and 32K token models alongside legacy text-davinci-003; includes chat history and pre-defined questions.
• Caches generated embeddings to a file so repeat queries skip recomputation.
• Provides a Hugging Face demo space for browser-based testing.

Caveats

• The author notes the documentation is outdated and maintenance time is scarce; the project is actively seeking contributors.
• Multiple PDF support, OCR, and local LLM backends (Falcon, Llama, etc.) are listed as upcoming, not shipped.
• Turbo models can return weaker answers when embedding similarity is low; the author recommends falling back to text-davinci-003 or GPT-4.

Verdict Worth a look if you want to study a minimal, pre-hype RAG architecture or need a lightweight PDF Q&A prototype. Skip it if you require production-grade multi-document pipelines, OCR, or fully offline operation.