What Happened
Retail discovery is shifting from showing more results to ranking the most useful ones. In a study of more than 21,500 consumers across 20 countries, 86% said whether a product works matters more than its brand. For grocery shopping and meal planning, 26% would let technology suggest options, but only 7% would let it decide and purchase [1].
That creates a clear role for retrieval-augmented generation (RAG): an assistant can explain and narrow choices, but its answers need current product, inventory and policy data. Vespa.ai positions real-time retrieval and ranking as a way to combine those signals. Reported deployments include discovery across more than 58 million Kleinanzeigen listings and a Vinted migration involving approximately one billion items; Vinted reported a 1.1% transaction increase and more than €3.5 million in gross merchandise value uplift [1]. Those outcomes are useful evidence, not a forecast for another retailer.
Why It Matters to Businesses
A fluent assistant cannot compensate for retrieving an unavailable, unsuitable or incorrectly priced item. Search quality now affects both the results customers browse and the evidence an assistant uses to make recommendations. Businesses should measure whether customers find suitable products with less effort, not just whether answers sound helpful.
Kimbodo Engineering Perspective
Use retrieval to establish facts and ranking to express business relevance. Keyword search remains useful for exact names and identifiers; vector search helps with descriptive or ambiguous requests. Combining them, then reranking against attributes such as availability, location and customer intent, is a practical starting point.
The tools play different roles. LlamaIndex, LangChain and Haystack can help orchestrate retrieval and answer generation. Weaviate, Pinecone, Qdrant, Milvus, Elasticsearch and Vespa are search or vector-infrastructure options to evaluate against the workload. We would not choose one on an embedding benchmark alone: update speed, filtering, ranking control, operational burden and total query cost matter just as much.
How We Would Implement It
- Define a canonical product record with identifiers, searchable text, structured attributes, inventory status, price, timestamps and visibility rules.
- Index text and embeddings together, with metadata filters for hard constraints. Stream inventory and price changes so retrieval does not rely on stale batch indexes.
- Retrieve candidates with keyword and vector queries, merge and deduplicate them, then rerank using relevance and permitted business signals. Keep paid-placement rules explicit rather than hiding them in the ranker.
- Ground assistant responses in retrieved product records. Show why an item matches, link to the item, and recheck price and availability before any purchase action.
- Evaluate with labeled queries and live experiments. Track suitable-result rate, zero-result rate, p95 latency, stale-item recommendations, conversion and returns—not answer quality alone.
Risks, Costs and Security
Embeddings, reranking and language-model calls add latency and cost; real-time indexing adds infrastructure complexity. Start with a measured search baseline, add each stage only when it improves outcomes, and use caching where freshness requirements permit.
Enforce access controls at retrieval time, minimize customer data in indexes and model prompts, and treat product descriptions and reviews as untrusted content that may contain prompt-injection instructions. Log retrieved identifiers and ranking decisions for investigation, while limiting retention of personal data. Keep recommendation and purchase permissions separate: the consumer study indicates substantially greater willingness to accept suggestions than autonomous purchasing [1].
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our RAG Development Services practice, or Estimate My RAG System.
