Updates Jan 6

PGrid

• Experimenting on search grounded models for doing both classification and as a method for dynamically adding items to the database with specs. This will be done by going through titles of listings, then using those models to create specs for the items assuming multiple models agree with each other. While this worked on popular items, there was a higher error rate for lesser popular models.
• Did some evaluating of Gemini advanced with deep research on its ability to generate articles and get specs or performance data. While it searches a lot of data, its accuracy still doesn’t seem 100% after manually checking its sources.
• Using gemini and claude, added a new article as well as updating the style of articles a little
• Fixed bug where a specific parsing error on ebay scraping can lead to infinite retries.
• New schema for SSDs and variants
• Created a frontend to help with classifications and do classification experiments. Current classification plan for SSDs is to have 2 models agreeing on the most important specs. One model being gemini with google search, the other is also gemini but with an input csv of SSD data.