TODO.md

export report

• export result in an excel file
• report the token length of the documents we feed
• if the result csv file exist, ask if run from start

Models support

• support MetaAI llama and Claude AI
• support Google API

chat options

1. context + paper + all questions
   • check mapreduce and stuff RetriveQA method
2. use GPT3 only if possible
3. load local index with embedding
4. for specific questions with chain of thought method

release

• markdown preparation code

test

• w/wo cheatsheet
• w/wo embedding
• extract content about the question
• small text/sentence embedding
• group question by relationship
• simply, 3Q per request?
• categorize the paper
• contain table or not

test parameters

• choose test set
• choose cheatsheet or not
• choose split question or not
• choose split content or not
• choose report or not
  • variation analysis
  • overall
• for section split content
  • test chuck overlap

Title screening

• use title to know if the paper is related

date file structure

• date file structure as an object

Use fine-tune method to memorize the cheatsheet

how many papers are bigger than context?

mark question as the same

Question consistancy check

• if plasma then active replication
• if naive then active replication
• if drug class contains then PI and INSTI
• if naive and no treated, should have number of individuals
• 4301 and 4302
• 4301 and 4101

Check No but AI get NA

Human x GPT-4 x Claude

if one of them get too much disagree in a paper. May need recheck human results.