Applicloud · Commerce engineering
Building an AI Product Advisor: Web Search or Catalog RAG?
For a store with 10,000 products and a technical sheet for each one, the decisive question is how reliably the assistant can retrieve the right evidence.
Technical blog · 16 November 2026 · Approximately 6 minutes
A customer asks: “Which sealant works outdoors, adheres to aluminium, and tolerates this temperature range?” A useful advisor must identify candidates, verify their specifications, distinguish variants, and explain its recommendation. Fluent text is the final step. The engineering challenge is connecting each claim to the correct product and document.
Two retrieval strategies, different responsibilities
With web search, the model searches the merchant’s domain and uses retrieved pages or documents to answer. This is a quick route to a prototype. OpenAI supports domain restrictions, but restricting the search does not establish that every technical sheet is discoverable or current.[1] Web search itself is a form of retrieval-augmented generation; the practical distinction is who controls the retrieval corpus and indexing process.
With catalog RAG, the merchant imports product data and technical sheets, extracts text, divides it into useful passages, and builds searchable indexes. Each passage retains its product ID, source URL, page and document version. At query time, the application retrieves evidence and supplies it to the language model. This adds ingestion and synchronization work, while making coverage and failed imports observable.
The economics: retrieval is only part of the bill
OpenAI’s standard web-search tool currently costs $10 per 1,000 search calls, plus model charges for retrieved content and generation.[2] At 10,000 customer questions per month, one search per question adds $100; three searches add $300. A question and a billable search call are different units, so multi-step product comparisons need separate measurement.
Embedding costs can be surprisingly small. Assume 10,000 documents
averaging 2,000 tokens: 20 million tokens. At $0.02 per million tokens
for text-embedding-3-small, embedding that text costs
about $0.40, before chunk overlap and added
metadata.[3]
This excludes PDF extraction, OCR, enrichment, storage, development
and operations. Both approaches still pay for the language model to
read evidence and write answers.
Build retrieval around product identity
For technical commerce, combine semantic search, keyword search and structured filters. Embeddings help match a customer’s intended use; keyword search preserves exact terminology and SKUs; database filters enforce constraints such as length, material and temperature. Normalize units before comparing numerical values, and preserve table headers during PDF extraction.
Keep product records and document passages linked through stable IDs. Retrieve candidate products, validate their constraints, and fetch supporting passages before generating the answer. Read price and availability from the live commerce API. Replace superseded document chunks during synchronization, and cite the source behind technical claims. A product-page conversation should pass its product ID into retrieval.
Local retrieval can reduce network-dependent search steps, but response speed still depends on embedding latency, reranking, model choice and answer length. Measure both first-token and completed-answer latency. When evidence is insufficient or conflicting, the advisor should ask a clarifying question or state the limitation.
Which commerce platforms already provide this?
“AI-powered commerce” covers several different capabilities: merchant copilots, generated descriptions, catalog search, and document-grounded customer advice. For this use case, evidence of document ingestion and retrieval matters more than the AI label. The following shortlist separates platform capabilities from third-party extensions.
| Platform / extension | Documented capability | What still needs validation |
|---|---|---|
|
Odoo 19 AI Agents |
Native agent sources include indexed PDFs, documents and Knowledge articles; agents can connect to Live Chat. [4] | Bulk ingestion, product-to-document mapping and catalog synchronization for the intended deployment. |
| Shopnix Enterprise | The vendor describes catalog/document ingestion, hourly synchronization and Qdrant-based hybrid retrieval. [5] | Supported PDF formats, OCR quality and contractual scope. This is a managed platform. |
|
WooCommerce My Observer |
A third-party RAG plugin with text-PDF ingestion and document-to-product linking. [6] | Throughput and operational maturity: its listing reports fewer than ten active installations. |
|
WooCommerce AskAny |
A third-party plugin indexing products, custom fields, PDFs and URLs, with hybrid retrieval and citations. [7] | Performance at 10,000 documents and reliable matching of technical sheets to product variants. |
|
Shopify Tech Guardian |
A third-party assistant using manuals, specification sheets and FAQs, with published catalog/PDF quotas. [8] | Document-page and conversation allowances, language quality and production references. |
Capabilities are based on vendor documentation, not independent benchmarks. Native functionality does not necessarily mean inclusion in every platform edition or subscription.
Document quotas deserve particular attention. Tech Guardian lists a $169/month plan with 10,000 products, 10,000 PDF pages and 450 monthly conversations. Its $399/month tier lists 30,000 PDF pages and 1,200 conversations.[8] Ten thousand three-page sheets already consume 30,000 pages. Product count alone is therefore a poor basis for estimating subscription cost.
Adjacent products also need careful classification. Spryker’s Storefront Assistant searches the catalog and CMS and accepts PDF attachments in conversations. Those features do not, by themselves, establish a persistent index of every product’s technical sheet. [9] Ask vendors to demonstrate that exact workflow.
Choose with an evaluation set
Before selecting a platform, assemble about 50 representative questions: exact SKU lookups, near-identical variants, numeric constraints, compatibility, conflicting document versions and questions the sources cannot answer. Evaluate retrieved evidence and product correctness separately from writing quality. Record citation accuracy, unsupported claims, latency and cost per successfully resolved question.
For an existing store, an external RAG service can preserve the commerce investment while adding technical advice through APIs and a chat widget. For a new deployment, native AI modules may reduce integration work. Either way, the durable engineering asset is a maintained relationship between products, specifications and evidence. That is what turns a conversational interface into a dependable product advisor.
