What is TF-IDF?
TF-IDF weighs terms by how important they are in a document and how rare they are across the corpus.
Interview Questions: Local TF-IDF + SVD Embeddings
Use these to explain local retrieval as a serious engineering tool, not a toy.
What does SVD add?
SVD compresses high-dimensional sparse TF-IDF vectors into lower-dimensional latent vectors that capture broader word-pattern structure.
When is TF-IDF strong?
Exact terms, IDs, error codes, product names, policy numbers, and domains where vocabulary overlap matters.
When is TF-IDF weak?
Paraphrase, multilingual search, deep semantics, context, negation, and vocabulary mismatch.
Why use local embeddings in a learning project?
They are transparent, private, cheap, dependency-light, and excellent for teaching retrieval mechanics.
What must be versioned?
Tokenizer, cleaning rules, vocabulary, IDF values, SVD projection, corpus version, and index release.
How can SVD hurt retrieval?
Too much compression can blur rare but important terms and make distinct documents look too similar.
How do you debug TF-IDF retrieval?
Inspect query tokens, missing tokens, matched terms, top weighted terms, scores, and whether cleaning removed important identifiers.
How does this fit with neural embeddings?
TF-IDF can be a baseline, fallback, exact-term engine, or hybrid partner with neural retrieval.
How do you evaluate it?
Use golden queries split by exact-term, paraphrase, stale-content, and domain-specific cases. Track Recall@k, MRR, and failure labels.