Skip to content
Coming soon

Retrieval that understands products, not just paragraphs.

A retrieval layer built on structured product meaning — so the systems asking about your catalog get grounded answers instead of plausible ones.

Powered by the Semantic Commerce Engine + Brand Knowledge Engine — part of the Semantic Commerce Layer™

What the retrieval sees

Two sources of truth. One grounded answer.

Generic RAG retrieves paragraphs. Semantic RAG retrieves this — structured product meaning plus verified brand knowledge:

// GET /rag/context — one retrieval, two sources
{
"catalog_context": { // Semantic Commerce Engine
"product": "a scented candle",
"brand": "a home fragrance brand",
"price": 20,
"availability": true,
"category": "home_goods › home_goods_decor",
"product_type": "candle",
"functional_intent": ["mood_setting", "room_brightening", "artisan_quality"], // what the product is for
"use_cases": ["home_styling", "office_decor", "seasonal_refresh"], // contexts the shopper asked for
"purchase_signals": ["cruelty_free", "phthalate_free"], // filters that decide the buy
"functional_attributes": { "clean_burning": true },
"materials": [], // empty — no evidence in the catalog
"search_synonyms": "+5 curated" // magnitude only — the curated terms are IP
},
"brand_context": { // Brand Knowledge Engine
"positioning": "verified brand record — story, promises, policies",
"shipping": "economy & expedited tiers · free over threshold · ships internationally",
"returns": "30-day window · store credit · exclusions listed",
"contact": "verified support channel",
"grounded": true // every field traces to published sources
}
}

If the catalog or the brand doesn't support a fact, the field comes back empty. The retrieval can only say what the merchant can prove — that's the difference between grounding and guessing.

Generic retrieval was built for documents. A catalog isn't a document.

Standard retrieval chops text into chunks and finds the passage that looks closest to the question. That works for articles. For products it fails in a specific way: it finds text that resembles the question while missing the product that actually satisfies it — because size, fit, compatibility and material aren't a matter of wording, they're facts.

The result is an answer that reads well and recommends the wrong item.

Built for teams putting AI in front of a catalog.

  • Merchants powering an assistant, search or recommendation over their own products
  • Platforms serving many catalogs that need consistent retrieval behavior
  • Teams whose generic RAG returns plausible but wrong products
  • Developers who want grounding, not guessing

What we need from you.

  1. 1.Your catalog (or your merchants').
  2. 2.Read access appropriate to the platform.
  3. 3.The kinds of questions your system needs to answer.

Access is credentialed, scoped and revocable — same rules as every NSOLVIA data surface.

What it does.

Retrieves over structured product meaning rather than loose text chunks.

Respects the facts that decide a purchase — variant, compatibility, availability — instead of treating them as words.

Grounds answers in the real catalog, so what comes back is a product that exists and fits.

Stays current as the catalog changes.

Rule we don't break: Verified, never invented.

What you get.

  • Retrieval grounded in structured product meaning
  • Consistent behavior across catalogs
  • Credentialed, scoped access
  • Documentation for the teams integrating it

Three steps.

Step 1

Connect the catalog

Step 2

It's enriched and indexed by meaning

Step 3

Your system queries it and gets grounded results

Plausible vs. correct.

Question
“will this fit a 14-inch laptop?”
Generic retrieval
returns the passage that talks about laptops → often the wrong product
Semantic retrieval
returns the product whose actual dimensions satisfy the constraint

What we claim: retrieval grounded in product facts.

What we don't claim: we don't control the system consuming the results.

Questions & answers

What teams ask.

Why not just use a standard RAG setup?

Because standard retrieval matches text, and a catalog decision depends on facts. Chunk-based retrieval will confidently return the passage that talks about the right topic while missing the product that satisfies the constraint — a fit, a size, a compatibility — which is exactly the failure that makes an assistant untrustworthy in commerce.

Does this replace my search?

It's the retrieval layer underneath systems that reason about your products — an assistant, a recommender, an internal tool. Whether it replaces or complements your existing search depends on what that search is doing today, which is part of what we scope with you.

Do I need the rest of NSOLVIA?

Retrieval is only as good as the product meaning underneath it, so it works best on an enriched catalog — the same structured data that powers feeds and the assistant. If your catalog is already enriched with us, this is a layer on top rather than a project.

What does integration involve?

Credentials, an interface and documentation. Your system queries it and gets grounded results — there's no index to build on your side, no embeddings to manage and no second copy of your catalog to keep in sync.

What happens when the catalog changes?

It stays current. New products, changed attributes and availability are reflected as the catalog changes, so retrieval doesn't quietly start returning products you no longer sell — which is the failure mode that makes teams stop trusting a system.

When is it available?

It's being connected now. Join the waitlist and you'll hear from us before it opens — and if you want the technical detail in the meantime, the technical overview is published.

Ground your AI in products, not paragraphs.

Semantic RAGsits inside NSOLVIA's interpretability stack — an open discipline, the framework that implements it, and the instruments that apply it.