Advanced RAG: the chatbot answers from your real catalog data
Advanced semantic search over products and documents: before replying, the chatbot retrieves information from your store, then forms the answer from that data.
What it actually does, step by step
What happens when a customer asks a question
Before answering, the chatbot goes looking. The question is matched against your knowledge base and the most relevant passages are pulled out: that product’s record, the paragraph of your return policy, the line in the price list. Only then does the language model write the answer, and it writes it from those passages. That is the difference between an assistant that knows where to look and one that answers from memory: the second one, when memory runs out, fills the gap with something plausible.
Semantic and keyword search together
Retrieval is hybrid: vector search, which captures meaning, runs alongside lexical search, which captures exact words. Both are needed. A customer typing "waterproof jacket for the mountains" will not use your catalog’s wording, and that is where the semantic side works; someone searching an item code, an abbreviation or a model number needs a literal match, which a vector alone blurs. Keeping both covers the shopper who describes what they want and the one who already knows its name.
What the chatbot can read
The knowledge base draws on three kinds of source. Your product catalog, synced from your Shopify or WooCommerce store, with records, variants and attributes. The documents you upload: PDF, Word, Excel, PowerPoint, CSV and plain text — price lists, manuals, terms of sale, spec sheets. And your own site pages, acquired by crawling. Long documents are split into passages before indexing, so retrieval returns the useful paragraph instead of the whole file.
Every answer knows where it came from
- Each indexed passage keeps its origin.
- The customer sees the page that says so.
- You can tell whether the answer or the document is at fault.
Answers can cite where they came from
Every indexed passage keeps its origin, so an answer can report the source it was drawn from. For the reader that is the difference between trusting and being able to check: faced with a warranty condition or a return window, the customer sees the page the information comes from. For you it is how you work out why the chatbot answered the way it did, and whether the problem is the answer or the document behind it.
Keeping the knowledge base current
The catalog updates with your store sync; documents update by handing over the corrected file. Re-indexing is idempotent: running it again updates the passages already there instead of piling up copies, so re-processing the same material neither inflates the store nor splits your answers in two. It is a boring property and that is why it matters — it means you can re-index whenever you are unsure, without first having to wonder whether you are about to do damage.
Where the text you index ends up
The embedding model follows the residency rule set for your store, just like the conversation model. The indexes stay in ShopBrain’s own database.
Where the data is processed
Embeddings are generated by a configurable provider, and that choice is bound by the data-residency rules set for your store: the text being indexed is your content and your customers’, and it follows the same constraint as the conversation model. The indexes live in ShopBrain’s own database, not in an external search service. For anyone who has to document how data is handled — all the more so under the AI Act — this question comes up early, and it is better to have the answer ready.
Can the chatbot make up product information?
Retrieval reduces the risk considerably, because the answer is built on data pulled from your catalog rather than on the model’s memory. No generative system removes the problem entirely, but there is a real difference between an assistant that starts from the actual product record and one with nothing to start from.
Which document formats can I upload?
PDF, Word (.docx), Excel (.xlsx), PowerPoint (.pptx), CSV and plain text files. Legacy Office formats are not supported: a .doc has to be saved as .docx before upload, and the error message says so when the file is rejected.
Do I have to re-index every time the catalog changes?
The catalog stays aligned through your store sync. Documents you uploaded by hand are yours to update, but the operation is repeatable with no side effects: re-indexing updates what is already there instead of duplicating it.
How is this different from a chatbot wired only to a model?
A chatbot without retrieval knows only what was in the model’s training data, where your catalog does not appear. It can talk about e-commerce in general and get every stock level, price and variant of yours wrong. Retrieval is what ties the conversation to your store.
Ready to answer in chat with your own data?
Free trial, no credit card. Set up in minutes.