velocitatem / ai-hackathon-ie-2023

🚀 submission V2

1. Our software uses python 3.11.5
2. pip install -r requirements.txt
3. python main.py [dataset] [output] --gpt4

• dataset - the directory with all the content to process
• output the name of the file to write to Ex: out.csv
• --gpt4 for better performance please use this option

ai-hackathon-ie-2023's Issues

Modify IO of the program

• Add arguments to the main program: dataset and output name
• Instead of getting the files by the Excel sheet, load all pdf and PDF from directory of interest
• If something is not found, put it as Nan
• Reformat data back to just a list for lists rather than 5 separate columns

Fix CSV Parser

Currently, the method for parsing the data into a CSV is not great, we should do something better than just appending strings, maybe the ideal would be to use df.to_csv()

Create a More Descriptive model for data

ai-hackathon-ie-2023/clai.py

Lines 7 to 17 in 1b303e9

	class Beta(BaseModel):
	Isin: str
	Issuer: str
	Ccy: str
	Underlying: List[str] = ["SX5E", "UKX", "SPX"]
	Strike: List[float]
	LaunchDate: str = "31.12.2021"
	FinalValDay: str = "31.12.2022"
	Maturity: str = "31.12.2023"
	Cap: int
	Barrier: int

Each of these items in the class have a 💩 description for now :(

Document page iteration

Instead of passing the whole document as one large string, it could be better to provide each page individually, extract the data, and then store it.

We are currently giving it all the data as a concatenated string, which causes issues with the tokens it takes to process the request (due to the string being too long), and some pages provide no relevant information, meaning that it could potentially cause an overload. One of the possibilities we thought about was to pass each page individually, helping the total length by keeping it short and extracting the data from each page with the possibility of discarding useless pages. However, one of the issues that might pop up by using this method is the model modifying correct data. The best solution for this is to get the accuracy scores for each individual page and make the most accurate results the final ones.

Improve Prompt Engineering

ai-hackathon-ie-2023/clai.py

Line 76 in 1b303e9

messages=[

🚧 WORK IN PROGRESS 🚧

Some resources:

• https://platform.openai.com/docs/guides/prompt-engineering
• https://developers.google.com/machine-learning/resources/prompt-eng

From some experience, it is best to show the model of the data we want to extract at the start and end of the prompt.