• We are using the CORD-19 (COVID-19 Open Research Dataset Challenge) dataset for testing.
• The data used in the POC is from the ./document_parses/pdf_json folder.

Note this is a 20GB huge dataset. >100GB when unzipped. Make sure you have enough space to hold everything.

sudo yum install -y jq cyrus-sasl cyrus-sasl-gssapi cyrus-sasl-plain

The script uses mongoimport together with jq to import the json files. mongoimport is included in the MongoDB Database Tools.

The following configuration should be replaced by yours:

• uri='mongodb+srv://<user>:<pwd>@vectorsearchpoc.wy1z6.mongodb.net/': The target MongoDB URI.
• db=covid: Database name where you want to store the dataset.
• coll=paper: Collection name where you want to store the dataset.
• json_source="<path_to_test_dataset>/document_parses/pdf_json/*.json": Where to find the json files.

./import.sh

Each json file is a big one has a lot of data included. We don't need all of them. Only the following fields are imported for this testing:

• paper_id: paper_id
• title: metadata.title
• text: body_text.text
• section: body_text.section

Note the body_text is an array. We unwinded the array so that each text snippet is inserted as one document.
This is an example of the final document structure:

{
  "_id": {
    "$oid": "6619439e4728952eae60aeb8"
  },
  "paper_id": "0000028b5cc154f68b8a269f6578f21e31f62977",
  "title": "\"Multi-faceted\" COVID-19: Russian experience",
  "text": "According to current live statistics at the time of editing this letter, Russia has been the third country in the world to be affected by COVID-19 with both new cases and death rates rising. It remains in a position of advantage due to the later onset of the viral spread within the country since the worldwide disease outbreak.",
  "section": "Editor"
}

• The application first load numSamples documents from the collection. The embedding in these documents will be used for later test.
• When a request arrives,
  • The application will randomly get 1 embedding from the previously loaded embeddings.
  • Use the embedding to do vector search.
  • The search result will return limit of numCandidates.
• The full document (without embedding because it's big and is usually useless to the client) will be returned to the client.

• NodeJS v18.
• Source code in /performance-test folder.
• Packages and tools:

cd performance-test
npm install
npm install -g pm2

To start:

pm2 start bin/www -i max --watch

To review logs:

pm2 log www

To stop:

pm2 stop www

To remove

pm2 delete www

Apache ab is used for testing

yum install httpd
ab -n 100 -c 1 http://localhost/vector-search