AI Book Recommendation Assistant

A recommendation the shop cannot fill costs more than it earns

A language model asked to recommend books will do it beautifully and will sometimes invent one. It will also cheerfully recommend a title that went out of print, quote a price from whenever its training data was collected, and suggest something the shop has never stocked. For browsing that is a curiosity. For a shop it is a customer who turns up for a book that is not there, which costs more goodwill than the recommendation earned.

Three retrievers doing different jobs

What makes this hybrid is not that it uses several techniques. It is that each one answers the question it is good at.

A recommender for taste

Classic recommendation algorithms handle what people who liked this also liked. This is a problem the field solved long before language models existed, and it stays solved by the method that solved it.

Semantic and vector search for description

A customer asking for something slow and sad set in a cold place is describing a book without naming anything a keyword search could match. That is the query type embeddings are for.

The live database for fact

Stock, price, availability. Never inferred, never remembered, always read at the moment of asking.

What the model is for

Understanding what was asked, and talking about it. A customer rarely states a query cleanly - they circle it, compare, change their mind mid-sentence, and want to argue about an author. The model handles the conversation and the interpretation, then hands the actual finding-of-books to the components that can be checked. It ends up doing the part it is best at and none of the parts where being fluent and being wrong look identical.

The two cheaper builds, and what each one costs later

Let the model answer from what it knows

A model trained on the internet does know these books, and that is the trap: it knows a title existed, not whether this shop has it, at this price, today. The answer arrives fluent and unfalsifiable, and the customer discovers the gap at the counter rather than in the conversation.

Fine-tune on the catalogue

Training the catalogue into the model looks like the durable version of the same idea. A catalogue changes daily and a fine-tune freezes it at the moment of training, so the model becomes confidently out of date with no way to tell a stale answer from a fresh one. Stock is a lookup, not a thing to learn.

Sounding authoritative is not the same as being retrieved

This is the same discipline as any knowledge base worth having, in a domain where the failure is easy to picture. The question is never whether a model sounds authoritative. It is whether the specific claim in front of you was retrieved from something authoritative, and whether the system knows the difference between the claims that were and the ones that were not.