FILTERData & productivity
Jev Sandwich for Text Extraction
Matt Hodgson@mhodgson2
This project utilizes a Jev-LLM-Jev workflow to extract text, reducing LLM costs and providing a confidence score for extractions.
SignalJev scores the answer.
Just used @typesafeai to make a “Jev sandwich” (jevwich?) for text extraction. Jev → LLM → Jev.
For the price of two Jev calls (~500ms and extremely cheap) you can halve your LLM cost and get a confidence signal you can trust.
1. Ask Jev where to look.
2. The LLM extracts from the subset Jev identified.
3. Jev scores the answer.
The same accuracy as feeding the LLM everything, way fewer LLM tokens, and every extraction comes back with a confidence score. In my testing, correct answers were > 0.95, and wrong answers were < 0.1.
Use the boatloads of cash you just saved to feed low confidence answers to a smarter model to increase your overall accuracy. This is going to be everywhere.