WANTED: Help my wife beat me in arguments!

Help me Build my Workflow
1

Hey all new here, new to AI and my wife needs your help! She has a terrible memory and is always getting her facts wrong during our arguments. I’m creating a fun workflow where I can upload a PDF document of the last 5 years of text messages 300 page 35mb) between us, extract the data while retaining timestamps, sender ID, and message content. I then query an agent to pull messages based on prompts such as “list all messages that talk about the kids school (keyword)”, “list all messages that have an aggressive tone (sentiment)”, or “summarize the conversation between x date and y date” (analysis)

I am familiar with parsing, chunking, embedding, vector databases, and context window limitations, but am overwhelmed by the sheer volume and variety of tools, pipelines and solutions.

Currently I am running everything locally (privacy) in docker on a MacBook Pro silicon, ollama chat model (llama3.2: 7b), default text extractor, ollama default embeddings, Postgres with pgvector.

Right now it’s doing a horrible job even extracting the text from the pdf and groups all message info such as date/time and sender, and then groups all the message content, and even splits messages further breaking the connection. Proud I’ve even got this far but just kinda stuck, and been seeing that RAG is challenging for these reasons.

I’m not a software engineer but I can navigate fairly well and write some basic code, but I can’t even tell if I’m over engineering or under engineering this project. I’m having a blast, my wife thinks It’s fun, I’m passionate about learning and would love to hear all your thoughts and advice.

Much love
-WarcrY

1 Like
2

This will never work because your wife already assumes she has won every such debate. You cannot improve upon those odds, so starting with the title, just concede (assuming you want her to continue being your wife).

I was also tempted to comment on the irony of using RAG, but I think you can figure out what could be said there, so I’ll leave that alone, at least for a week or so.

Tuning and constraining the output from AI is an artform (not too different at times from arguing with a spouse), so just keep trying different prompts with different instructions until you figure out the right set of things to get the outcome you want… or maybe just send it flowers and see if that works.

The best strategy (from what I’ve observed) for discovering how to make things work, is to follow several tutorials like this one and adapt as best you can to your own problem domain. Ultimately, though, it may become necessary to engage a counselor. If you’re not already using Discord, start here.

3 Likes
3

:rofl:Thanks for the laugh! I’ll definitely take a look at that blog tutorial. After looking around the forum I noticed people uploading their workflow diagrams. Maybe I should do that to give people a better understanding of what I’m working with.

Again, thanks for the laugh, and your experience, strength, and hope.

Cheers!