Medical Search Engine with SPLADE + Sentence Transformers in Python
In this video, we'll build a search engine for the medical field using hybrid search with NLP information retrieval models. We use hybrid search with sentence transformers and SPLADE for medical quesiton-answering. By using hybrid search we're able to search using both dense and sparse vectors. This allows us to cover semantics with the dense vectors, and features like exact matching and keyword search with the sparse vectors. For the sparse vectors we use SPLADE. SPLADE is the first sparse embedding method to outperform BM25 across a variety of tasks. It's an incredibly powerful technique that enables the typical sparse search advantages while also enabling learning term expansion to help minimize the vocabulary mismatch problem. The demo we work through here uses SPLADE and a sentence transformer model trained on MS-MARCO. These are all implemented via Hugging Face transformers. Finally, for the search component we use the Pinecone vector database. The only vector DB at the time of writing that natively supports SPLADE vectors. 🔗 Code notebook: https://github.com/pinecone-io/examples/blob/master/learn/search/hybrid-search/medical-qa/pubmed-splade.ipynb 🎙️ AI Dev Studio: https://aurelio.ai/ 🎉 Subscribe for Article and Video Updates! https://jamescalam.medium.com/subscribe https://medium.com/@jamescalam/membership 👾 Discord: https://discord.gg/c5QtDB9RAP 00:00 Hybrid search for medical field 00:18 Hybrid search process 02:42 Prerequisites and Installs 03:26 Pubmed QA data preprocessing step 08:25 Creating dense vectors with sentence-transformers 10:30 Creating sparse vector embeddings with SPLADE 18:12 Preparing sparse-dense format for Pinecone 21:02 Creating the Pinecone sparse-dense index 24:25 Making hybrid search queries 29:59 Final thoughts on sparse-dense with SPLADE #artificialintelligence #nlp #naturallanguageprocessing #machinelearning #searchengine
