An experiment driven Search Engine Project, developed to index and retrieve best documents given a query using ensemble of models.

• Filter Models
  • BM25
  • TF-IDF
• Voter Models
  • MPNET
  • RoBERTa

• Phase 1
  • Data Analysis & Pipeline
  • Model Pipeline
  • Evaluation Pipeline
• Phase 2
  • BM25 Model + MPNet Model
  • Hyperparameter tuning
  • Ensemble Pipeline
• Phase 3
  • RoBERTa Model
  • Ensemble enhancement
  • Experimentation

• Finetune ColBERT
• Implement Clustering of docs

Note : The project was tested on linux and MacOS. (Windows has dependency issues, refer Troubleshooting)

1. Clone repository

   $ git clone https://github.com/TF4ces/TF4ces-search-engine.git

2. Setup Environment repository

   $ python3 -m venv venv
   $ source venv/bin/activate                [LINUX/MAC]
   $ .\venv\Scripts\activate                 [WINDOWS]
   $ pip install -r src/requirements.txt 

3. Download pre-loaded embeddings to this path: ./dataset/embeddings_test from GDrive

   Note: To generate embeddings from scratch run./tests/test_evaluate_model.py script setting MODEL to all-mpnet-base-v2, all-roberta-large-v1 individually twice.

   WARNING: use a GPU machine and it is expected to take 1hr to generate.

4. Run TF4ces Search Engine [install jupyter by $pip install jupyter notebook and to run $jupyter notebook]

   1. Run Eval Pipeline from ./tests/notebooks/TF4ces_Search_Eval.ipynb ipynb notebook.
   2. Run prediction Demo Pipeline from ./tests/notebooks/TF4ces_Search_Demo.ipynb ipynb notebook.

1. Windows Systems are seen to have issue while reading data with ir-datasets==0.4.1

   For windows the doc.iter might throw decoding error while reading tsv file, You would need to change the encoding in source files of dependency as per this issue.

   Issue : allenai/ir_datasets#208 (comment)