AI Document Workbench

Interact with your documents using reliable AI.

Welcome to the AI Document Workbench

This tool allows you to "chat" with your documents. Instead of relying solely on general knowledge, this AI looks up specific information from the files and links you provide to give you accurate, sourced answers.


Technical Decisions & Architecture

Embedding Optimization

We initially experimented with the local sentence-transformers/all-MiniLM-L6-v2 model.

However, to optimize cloud deployment and minimize server compute load, we opted for an API-based embedding approach using OpenAI.

LLM-as-a-Judge

We conducted a Proof of Concept using RAGAS for benchmarking.

To keep API token load low in production, we replaced the RAGAS framework with a lightweight, built-in "LLM-as-a-judge" to verify accuracy on the fly.

Docker Containerization

The application is fully containerized using Docker. This ensures consistent performance and streamlines the CI/CD pipeline.

Azure Deployment

The live site is hosted on Microsoft Azure App Service. The Docker image is pushed to a private Azure Container Registry (ACR).


How it Works

This application uses a technique called Retrieval-Augmented Generation (RAG). Here is the process:

  1. Ingestion: When you upload a PDF or link, we split the text into small "chunks".
  2. Embedding: We convert these chunks into numerical vectors (lists of numbers) that represent their meaning.
  3. Retrieval: When you ask a question, we search the database for the chunks most similar to your query.
  4. Generation: We send your question + the relevant chunks to the AI model, which writes an answer based only on that information.

Knowledge Base Status

⏳ Your Uploads (Temporary):

Any document you add via the "Add Knowledge" tab is considered temporary. To keep the database clean and efficient, user uploads are automatically deleted 1 hour after creation.