Back to projects

01 / AI Engineering / RAG

RAG PDF Reader

A multi-service RAG application that turns PDFs into searchable context and generates answers grounded in retrieved document content.

Overview

A multi-service retrieval-augmented generation application for working with PDF documents. It connects ingestion, chunking, embeddings, vector storage and retrieval to an LLM answer-generation stage.

Problem

Finding relevant information in long PDFs requires more than keyword matching. The application makes document content retrievable as context, so a language model can answer questions using relevant passages.

Architecture

Documents move through an ingestion and indexing pipeline. At question time, relevant content is retrieved from Qdrant and supplied as context for the LLM answer.

  1. PDFDocument input
  2. ChunkingSplit document content
  3. EmbeddingsRepresent the chunks
  4. QdrantStore document vectors
  5. RetrievalFind relevant context
  6. LLM answerGenerate with context

Implementation

  • Ingest PDF content, split it into chunks and create embeddings for vector storage.
  • Store vectors in Qdrant and retrieve relevant document content for context-aware answers.
  • Bring together Python, Django, LlamaIndex and Google GenAI in a multi-service application, with Docker and Inngest in the supporting stack.

Challenges

Chunk boundaries, retrieval relevance and the hand-off between services are central engineering concerns. A retrieved passage can be relevant without being sufficient to answer a question; answer quality therefore depends on the full pipeline.

Technology

  • Python
  • Django
  • Qdrant
  • LlamaIndex
  • Google GenAI
  • Docker
  • Inngest

Outcome

A connected PDF-to-answer workflow spanning ingestion, vector indexing, retrieval and generation. No retrieval-quality or latency benchmark is claimed.

What I learned

A RAG application is as much a data-pipeline problem as a prompting problem. Chunking, embeddings and retrieval all shape the context the model receives; each stage needs to be understandable and independently assessable.

Explore the other projects