Embedding Search and Semantic Similarity

• Embed documents and queries into a shared space
• Rank by cosine similarity with correct math and indexing
• Combine lexical + semantic signals for hybrid ranking

• Users search by concept; keyword search fails on paraphrase ('how do I change my pwd?' vs 'password reset').
• Embeddings are cheap to serve: precompute doc vectors, embed the query once, retrieve by index.
• Hybrid retrieval measurably beats either alone on real corpora.
• Semantic similarity powers more than search: dedup, clustering, question-answering context.

The title catalogue finds exact matches fast; the subject index finds 'books about this idea' whose titles never overlap. The best library uses both and merges the results — that's hybrid search.

• Embed legal docs (chunked) with a sentence/retriever model into dense vectors.
• Normalize and store in a vector index (faiss, hnswlib, pgvector, Milvus).
• At query time embed the query with the SAME model — model parity is non-negotiable.
• Retrieve top-k by cosine (inner product on normalized) with ANN for scale.
• Hybrid: BM25 lexical scores + semantic scores → rank fusion (RRF or learned blend).

• Customer support RAG retrieves resolution docs by semantic meaning ahead of LLM answer.
• Product search blends embedding relevance with a boosted exact-match for SKU codes.
• A knowledge base dedupes near-duplicate articles via similarity clusters.

• Embed at chunk size that matches the query's granularity and the answer length.
• Use the same retriever model at index and query time; freeze version changes behind a rebuild.
• Normalize vectors; pick one distance metric and keep the index consistent with the metric.
• Build a hybrid with BM25 and tune fusion weights on a labeled relevance set.

• Mixing embedding-model versions across index and query (silent distribution gap).
• Retrieving by raw cosine on unnormalized vectors — magnitude dominates similarity.
• Treating recall@10 on synthetic eval as production relevance — judgments need real queries.
• Over-chunking legal docs into units smaller than the answer, losing the answer across chunks.

• For a broken match, print the top-3 nearest neighbors' text — hard negative inspection is fastest.
• If short queries underperform, embed with padding or use the query-aggregation convention of your model.
• ANN recall < 100% — tune ef/candidates trade-off; monitor recall on a quality sample.
• Compare a lexical-only baseline: metric lift proves the semantic layer adds value.