•   Davit Shavit https://github.com/davis0011
•   Noam Kent https://github.com/Kentno

•   Take Parquet files from bucket
•   Use slightly modified version of assignment 3 gcp code to make indeciese
•   Create a dict of metadata on each document, ex. full title with no tokenizing/stopword removal
•   Transfer all data to an instance to run the retvival engine

All search functions other than "search" are according to the requirements.

•   Search: implemented using a combination of BM25 on the body text, a count of the number of words in the title and a binary decision on anchor text.

• The scores are weighted 2-4-1, which was reached by testing severeal times on different parts of the training set.

•   The main code block is the calculation of BM25 scores for each document in a numpy matrix, there is an explination of this in the report. See the graphic below: