Keyword search breaks the moment someone describes what they want in their own words. This Specialization covers semantic search: representing text as vectors so retrieval works on meaning. It runs from raw text through embeddings, indexing, hybrid retrieval, and evaluation, and underpins most RAG systems.
You prepare and tokenize text, generate dense embeddings with Sentence Transformers, and compare vectors with similarity and distance metrics. You build FAISS indexes, fuse dense retrieval with BM25 keyword search, add metadata filtering, then measure quality on a labeled set and improve it with chunking and reranking.
By the end of this Specialization, you will be able to:
• Prepare, tokenize, and numerically represent text for retrieval.
• Generate and persist dense embeddings with Sentence Transformers.
• Compare vectors using cosine similarity, dot product, and distance.
• Build FAISS indexes and reusable retrieval pipelines.
• Fuse dense and sparse rankings with metadata-aware filtering.
• Measure quality with Precision@k, Recall@k, MRR, and reranking.
This Specialization suits machine learning engineers, AI engineers, backend developers, data scientists, and search engineers adding retrieval to their products, plus developers preparing for RAG work. It assumes basic Python and no NLP background.
Enroll now to build a semantic search system you can measure and improve.
Übungsprojekt
Across the Specialization, learners complete hands-on projects that build one semantic search system from raw text to a measured, tuned pipeline. Learners clean and tokenize a product dataset, encode it with Sentence Transformers, and persist the embeddings so inference is not repeated. Learners then build a brute-force search, replace it with FAISS indexes, add a BM25 keyword baseline, and fuse the two into hybrid retrieval with metadata filtering, packaged as a reusable retrieval class. In the final project, learners create a labeled evaluation set, measure Precision@k, Recall@k, and Mean Reciprocal Rank, produce an error-analysis report, and improve ranking through chunking, query expansion, and cross-encoder reranking. Each project produces a concrete artifact learners can add to a professional portfolio.
