Proof of Concept: Small Model Accurate CSV Parsing with Auto Debugging Code Execution Workflow
Using Ollama models and a CSV data file, it was troublesome to accurately parse for values. (The LLMs would almost get it right, but skip values present in a search, except for under certain prompts and conditions.)
So, I started wondering, would it be easier to generate code to parse the CSV directly using the LLM, perhaps with an auto-debugging workflow, so that general user queries can be translated into exact data parsing algorithms on the fly?
After much hard debugging, and learning a lot about experimental state machine driven execution of node based workflows, I was able to get a successful result:
So here is the CSV Data:
And here is the output of the Workflow:
As you can see, the user prompt directing the process produced a successful result, capturing all the values and formatting the response in a way the user can easily see where the unknown values resided in the data.