Use the JITR Bot to generate context-aware responses¶

Access this AI accelerator on GitHub

Retrieval Augmented Generation (RAG) has become an industry standard method for interfacing with large language models by making them 'context-aware'. However, there are a number of situations where a text generation problem is not solved by interacting with large vector database containing many documents. These problems require context but where the context is not known before query time and is often unrelated to existing vector stores. Usually, they are questions about single documents where desirable behavior is to allow the document to be specified at runtime.

One application that does this fairly well is DataChad, built by DataRobot's own Gustav von Zitzewitz. Datachad works fine for its purpose as a localized web application but it doesn't generalize. In other words, there is not a good way to interact with the application without opening a browser, uploading whatever files you want to analyze, and hitting a run button.

Rather than follow the standard RAG approach of querying an existing vector store, this accelerator creates a deployment that accepts a file as an argument so that it can provide context-aware answers 'on the fly'. DataRobot calls this approach "Just In Time Retrieval", or JITR for short. DataRobot created a Slackbot that uses this deployment as the backend to answer questions when a user uploads a PDF, called the "JITR Bot".