Key Concepts
- 1Retrieval is how the agent finds candidate products before it reasons — garbage retrieval means the LLM reasons brilliantly over the wrong 20 items
- 2Keyword/lexical search (BM25) is precise for exact terms but blind to synonyms and intent; vector/semantic search captures meaning ('cozy' -> fleece, flannel) but can drift off-topic
- 3Hybrid retrieval — combining lexical and semantic with a re-ranker — is the practical default; it catches both exact-match SKUs and intent-level matches
- 4RAG (retrieval-augmented generation) grounds the agent in your actual catalog so it recommends real, in-stock SKUs instead of hallucinating plausible-sounding products
- 5Filters are non-negotiable in commerce: price, availability, size, and region must be hard constraints applied at retrieval time, not hoped-for in the LLM's output
- 6Embedding quality depends on what you embed: rich product text (title + attributes + reviews) retrieves far better than a bare title; catalog data quality is upstream of everything
- 7Latency budget: retrieval + re-rank must fit inside the conversational response window, so PMs trade off index size, top-k, and re-ranker depth against speed
Why It Matters for PMs
Recommendation ceiling: The agent can never recommend a product it didn't retrieve. Retrieval quality is a hard upper bound on the entire experience — no prompt cleverness fixes bad candidates.
Hallucination control: RAG grounding is the primary defense against the agent inventing products or prices, which in commerce is a trust and legal problem, not just a quirk.
Catalog ROI: Retrieval performance forces investment in structured catalog data. PMs who fund catalog enrichment unlock compounding gains across search, recommendations, and merchandising.
Test Your Understanding
Question 1 (Multiple Choice)
Why is retrieval quality a hard ceiling on agent recommendations?
Question 2 (Multiple Choice)
What is the main advantage of hybrid (lexical + semantic) retrieval over pure semantic search?
Question 3 (Multiple Choice)
In commerce RAG, price and availability filters should be applied...
Question 4 (Short Answer)
Your agent keeps recommending out-of-stock or out-of-budget items. Where in the retrieval pipeline would you look first, and why?
Think about your answer before revealing the explanation.
Question 5 (Short Answer)
What's the tradeoff a PM manages between retrieval recall (top-k size) and conversational latency?
Think about your answer before revealing the explanation.