Conversations with the Past

How ChatGPT and other LLMs can connect us with otherwise lost documents in a way no product can hope to

While AI has changed how we look at the future, there are also ways to use it to better understand the past. We read old documents through what has been recorded and transcribed into digital text. If it isn't convenient to read and search, we largely ignore old documents. Converting anything analog to digital takes a lot of time and effort. AIs provide us with ways to transcribe and refine these documents and converse with them.

Midjourney - Prompt - castle and bailey ink scroll layout drawing —v 4 --niji --ar 2:1 --q 2

I have my bias that AI is an exciting topic. Still, I think this "talk to documents" capability gets around the hype and shows off something very useful. Everyone has old files or old documents that can't easily be digitized. No one has time to do much with them. Why not offload the effort on an AI and get it off your old bucket list? What's possible here is easy enough to set up and better than most paid software available for individuals.

I found the capability for AI to assist with documents as I was reading through the documentation for an AI tool called LangChain. While they have a conventional search for their documentation, they also have an AI-powered chatbot whom you can ask for directions and explanations and even help brainstorm use cases that aren't directly documented. Talking to it was like conversing with a textbook and getting back a clear and well-documented description, including sources.

I wanted to talk to every book or presentation I read to ask questions about the content. (I'm sure this is a Kindle feature in the making.) It was intoxicating but, unfortunately, sometimes impossible to set up. ChatGPT has a limited length of text that can be input, and that limits your conversations to those with painfully short stories.

To help solve this concern, LangChain, and tools like it, provide ways to work around those limitations and enable these incredible experiences. I've used this to help ask multiple questions about a presentation I didn't understand. I could ask follow-up questions as I better understood. It's like putting on your own Q&A with the speaker.

As a good millennial, I, of course, adopted these ideas as a side hustle and built an app following this concept. I've been playing an old version of Dungeons of Dragons (Advanced D&D 1st Edition) that came out in the late 70s. To play, we flip through poorly converted PDFs that contain the confusing ruleset of that time. It's a huge undertaking just to play the game. It would be great to ask the rules in plain English how to play.

Taking what I could get out of the PDFs, I set ChatGPT to fix much of the text. This gave me something much more searchable for me and the AI. Then I set LangChain loose on it! The result, as depicted below, is a fantastic experience where I can converse with these old rule books and get referred to pages containing the result I've been given back.

Nerdy Note: Part of the magic sauce that makes this super powerful is breaking the PDF into pieces and uploading chunks with metadata that help LangChain send the most relevant data to the LLM. This makes the prompt much more concise and less likely for the AI to get confused about what it has been sent.

There are a ton of great use cases for this toolchain. Many different types of documents with a diverse set of needs. I'm much more excited about the individual empowerment of this sort of tool than I am any individual product offering. Sure, Plugins and the likes are great. Yes, there will be updates to send more data through ChatGPT or other LLMs. But this solution can better fit your need than anyone offering a product today!