Module 4 · Project 2: Smart FAQ Search (semantic search with sentence embeddings)

OBJECTIVE: build a search box where a customer types a question in their own words and gets the 3 best answers from a FAQ — matching by meaning, not exact words — and learn sentence embeddings and semantic search while building it. WHAT WE COVER (in project order): • Defining the problem and search scores (Recall@k, MRR) • Building a small FAQ knowledge base and a labelled test set of questions • Keyword search baseline (BM25 / TF-IDF) • Sentence embeddings and cosine similarity • Building the search with a vector index (FAISS) • Improving results with a re-ranker • Evaluation, a 'no good answer' threshold, and comparison with Project 1 • Deployment: Gradio search app AFTER THIS MODULE YOU CAN: • Build and evaluate a semantic search system without training a model • Explain embeddings, vector search and re-ranking • Build the 'retrieval' half of RAG (the generation half comes in the LLM module) DATASET: a FAQ you create (one short answer for each of the 77 Banking77 intents) + Banking77 test questions as the test set (we already know the correct FAQ for each). TOOLS: sentence-transformers, FAISS, rank-bm25.