Where this book stops, and what is beyond it
The eleven chapters build one search engine, from exact word matching to a trained two-tower retriever. AI-Powered Search goes further, into knowledge graphs, personalisation, click models, active learning, LLMs and multimodal search. Two of those ideas are demonstrated below; the rest are named honestly, with a link to the closest thing this app does have. Nothing here is claimed to be production search.
Two frontier ideas you can run
An answer, with its sources, on top of the same results
Retrieval-augmented generation returns prose, not a list. The loop is: retrieve, write only from what was retrieved, cite it, then check the claims back against the sources. This demo writes the sentences from listing fields with fixed templates, so the loop is visible without a language model. A real system replaces the writer, not the retrieve–cite–verify contract.
No listing shares a word with “sofa”, and nothing is close in meaning. There is no evidence to write an answer from. Try a product word such as “sofa” or “camera”.
The checker tokenises each sentence and looks for its content words in the retrieved listings. A word that appears in none of them turns red: that is the kind of unsupported claim a stronger verifier would stop before the answer reaches a person. This is a token check, not entailment: it catches new claims, not a misused one. Retrieval bounds what can be said; the checker bounds what is allowed to stand.
The same results, re-ranked for one person
Personalisation adds a boost that depends on the searcher, not the query. The model below is deliberately transparent: a listing gains points when its category and city match the persona. A production system learns those weights from signals; here you can see exactly what each point is worth.
Looks for furniture and home goods, and buys locally.
0 of 0 listings move. At weight 0 this is chapter 9's order.Trade-off, stated plainly: this makes the list better for one persona and worse for everyone else, and a profile can narrow what a person is shown. That is why the book treats personalisation as a boost to be measured, not a replacement for retrieval.
What is not built here
A reviewer should be able to see the boundary at a glance. Each row names a book chapter, what the app actually does, and where to look.
- Book chapter 5 · knowledge graphsKnowledge graphsA flat, hand-written table of brands, models and product types. There is no graph, no extraction and no traversal.Open the entity table →
- Book chapter 9 · personalised searchPersonalisationA transparent category-and-city boost, shown above. There is no learned user embedding and no collaborative filtering.Try the boost demo ↑
- Book chapter 11 · click modelsClick modelsPosition bias and inverse propensity weighting, worked through on one page. There is no dynamic Bayesian network and no confidence prior.Open position bias →
- Book chapter 12 · active learningActive learning and A/B testsOffline metrics only. There is no online experiment, no exploration policy and no statistical test.Open the scorecard →
- Book chapter 13 · dense vectors at scaleApproximate nearest neighboursExact search over the full WANDS catalog (about 43,000 products). HNSW, quantization and recall-versus-latency are named but not built.Open hybrid search →
- Book chapters 14-15 · LLMs and foundation modelsAnswering, grounding and multimodal searchThe grounded-answer demo above, with a template writer instead of a language model. No fine-tuning, no generative or multimodal search.Try the answer demo ↑
What this app does not claim
The catalog is the WANDSThe public product-search dataset this app searches. It provides the catalog (42,994 products) and the human relevance grades (Exact, Partial, Irrelevant) for the evaluated queries. dataset (Wayfair ANnotation Dataset, MIT licence, github.com/wayfair/WANDS): 42,994 products and human grades (Exact, Partial, Irrelevant). Price and date are deterministic synthetic overlay columns because WANDS has neither. The numbers reproduce on this machine and are not a production experiment. The app teaches the mechanisms; the book teaches the mechanisms and the scale. Use the app to see how a ranking is built and measured, then read the book for the systems that serve real traffic.