Retrieval-augmented generation (RAG) is a method where an AI system first searches a set of your own documents, such as product data, policies and help articles, then writes an answer based on what it found. For e-commerce brands, RAG powers shopping assistants, support bots and internal tools that answer from current catalogue and policy data instead of the model’s general memory.
The idea came from a 2020 research paper by Patrick Lewis and colleagues at Facebook AI Research and partner universities. It is now the most common pattern for business AI assistants, because it keeps answers tied to sources the business controls and updates.
How does RAG work, step by step?
A RAG system has two halves. The first prepares your content: documents are split into chunks, each chunk is converted into a numerical representation called an embedding, and those embeddings go into a search index, often a vector database. The second half answers questions: the user’s question is embedded, the closest chunks are retrieved and the language model writes an answer using them.
- Ingest: collect product feeds, policy pages, help articles and FAQs.
- Chunk: split them into passages of a few hundred words, keeping product records whole.
- Embed and index: store passages in a vector index, often with keyword search alongside.
- Retrieve: find the most relevant passages for each question.
- Generate: the model writes an answer citing or using only those passages.
- Update: re-index when products, prices or policies change.
Many platforms package these steps, from cloud services such as Azure AI Search, Google Vertex AI Search and Amazon Bedrock Knowledge Bases to e-commerce support tools with built-in knowledge bases. Custom builds use frameworks such as LangChain or LlamaIndex.
Which e-commerce use cases suit RAG?
Shopping assistant
A customer asks, “Which of your running shoes work for wide feet under a certain budget?” A RAG assistant retrieves products whose data mentions wide fit and filters by price, then explains the options. The quality depends heavily on whether fit information exists in the product data. If it does not, no model can invent it reliably.
Support and policy answers
Return windows, shipping zones, warranty terms and payment options change. A RAG support bot answers from the current policy page, so updating the page updates the bot. That is a major advantage over fine-tuning, where the model would need retraining for every policy change.
Internal tools
- Customer service agents searching policies and past resolutions during a live chat.
- Merchandisers asking which products lack size charts or care instructions.
- New staff onboarding with answers drawn from internal process documents.
- Buyers querying supplier specifications across hundreds of PDFs.
Content support
Writers can generate product comparison drafts or buying guides grounded in actual product attributes, reducing invented specifications. An editor still checks every claim, but the starting point is anchored in data.
What makes a RAG system answer badly?
Most RAG failures are retrieval failures. If the right passage is not found, the model either answers vaguely or, worse, fills gaps with general knowledge. Common causes are poor chunking, outdated indexes and product data that lacks the details customers ask about.
- Stale index: prices or stock changed but the index was not refreshed.
- Missing data: attributes such as fit, material or compatibility are absent from the feed.
- Bad chunking: a policy table split across chunks loses its meaning.
- Synonym gaps: customers say “trainers”, product data says “sneakers”.
- No refusal rule: the model answers even when retrieval found nothing relevant.
Hybrid search, which combines keyword and vector search, fixes many synonym and product code problems. A clear instruction to say “I don’t have that information” when sources are missing prevents most invented answers.
How do you test a RAG assistant before launch?
- Collect 100 real customer questions from chat logs and emails.
- Write the correct answer and the source page for each.
- Measure retrieval: did the system find the right source in its top results?
- Measure answers: correct, partly correct, wrong or appropriately refused.
- Include tricky cases: out-of-stock items, discontinued products, other languages.
- Retest after every major catalogue or policy change.
Separate the two scores. If retrieval is good but answers are wrong, fix the prompt or model. If retrieval is poor, no prompt will help; fix the data and the index.
Language deserves its own test round. A store selling in Turkish and English should check that a Turkish question retrieves English-only product data correctly, or better, that key attributes exist in both languages. Embedding models differ in multilingual quality, and a system that scores well in English can quietly miss half the Turkish questions.
What about privacy and safety?
Keep customer personal data out of the shared index. Order lookups should go through a secure, authenticated API call, not through documents stored in a vector database. Also watch for prompt injection: if product reviews or supplier documents are indexed, text inside them could try to instruct the model. Treat indexed user-generated content as untrusted.
How Alien Road applies RAG for e-commerce clients
A typical engagement starts with a question audit: we pull several months of chat and email questions, group them by topic and check whether the client’s product data and policy pages actually contain the answers. The gaps found here often matter more than any technology choice, because missing fit or compatibility data blocks the assistant from day one.
Setup usually uses an existing platform where it fits the client’s stack, with hybrid search, a refusal rule and a scheduled re-index tied to the product feed. We build the 100-question test set with the client’s support team. Measurement tracks retrieval accuracy, answer accuracy, handoff rate to human agents and customer satisfaction on bot conversations, reported monthly with examples of failed answers and the fixes made.
Where to start this week
Export 100 recent customer questions and, for each one, find the page on your site that answers it. Mark the ones with no answer anywhere.
- Collect questions from chat, email and marketplace messages.
- Match each to a source page or mark it as missing.
- Fix the top five missing answers on the site itself.
- Check whether your product feed includes the attributes customers ask about.
This exercise improves your site immediately, and it is exactly the test set a RAG system will need later.
Keep reading: AI Optimization · AI Optimization