Semantic Search and Agentic Search in B2B: When Vector Embeddings, When LLM Agents?

Semantic search is no longer a research topic in eCommerce. Vector embeddings run in production inside Algolia NeuralSearch, Constructor, Coveo and Bloomreach Discovery; native pipelines can be built on Snowflake with Cortex. In parallel, "Agentic Search" has arrived as a new stage — the LLM agent as advisor, querying multiple sources, asking follow-up questions, guiding multi-step conversations. In a B2B context this does not mean: everything must become AI. It means: you now have three maturity stages — keyword plus synonyms, semantic with embeddings, agentic with LLM — and you have to decide which fits which search occasion. This article closes our B2B search series with the question that comes up most often in architecture workshops: what does AI actually contribute to search — and what does it cost?

Vector embeddings turn search queries and products into mathematical vectors whose proximity reflects semantic similarity. A search for "slip-resistant protective gloves for oil environments" then also surfaces products titled "nitrile gloves, grip coating, oil-grip" — without anyone maintaining synonyms by hand.

B2C studies regularly report double-digit conversion lifts from semantic search. Those numbers should not be carried over to B2B unchecked — search occasions are different, the catalog is smaller, click behavior is different. Where embeddings do deliver measurably in B2B, it is in three constellations: long natural-language queries like "lubricant for high-temperature bearings up to 250 degrees", queries with descriptive vocabulary instead of SKUs, and catalogs with pronounced synonym variety across DIN/ISO designations, brand names and industry jargon.

Where embeddings do not deliver is just as clear: in pure SKU and part-number search. Pure dense vector search degrades on exact product IDs, SKUs, measurement designations — precisely the behavior that accounts for the larger share of B2B search. The productive consequence is hybrid search: keyword (BM25) and vector run in parallel, the result sets are merged and ranked together. Algolia NeuralSearch does this in a single API; Elastic, OpenSearch and Snowflake's native Cortex Search service do too. Anyone thinking about semantic search in a B2B setup is in practice thinking about hybrid search.

Agentic search is not the next version of semantic. It is a different interaction model. A classic search system answers a query with a result list. An agentic search assistant breaks a request into multiple subtasks, calls tools — product catalog, availability, configurator, ERP — and runs a dialogue until the request is resolved. Amazon has had this pattern in production with Rufus since February 2024 in B2C.

In B2B this shifts search toward configurator logic. A request like "We are building a conveyor system for bulk material up to 80 tons per hour, outdoor use, FDA-compliant — which bearings fit?" cannot be answered with a meaningful result list from a classic index. An agent that clarifies the application parameters, then filters the catalog, checks datasheets, adds cross-sell components and finally presents three candidates with reasoning is closer to what buyers know from personal sales conversations. This is exactly where cross-sell potential sits: not in "customers also bought", but in context-driven combinations — matching lubricants for the bearing recommendation, matching fasteners for the construction.

The architecture behind this is not trivial. A productive agent needs tool use (API calls to catalog, pricing, inventory), multi-turn state (the system remembers what has already been clarified), and an abort logic when the agent drifts into a hallucination zone. In B2B there is an additional requirement that the B2C discovery context does not have: the agent must not invent a recommended component. It must exist in the catalog, the customer must be allowed to order it, and the price must be the contract price.

For implementing semantic and agentic in B2B there are two realistic paths — rarely either-or, often a deliberate split.

Path A: Native search platform. Algolia NeuralSearch delivers hybrid search from one API; Constructor and Coveo bring similar capabilities. Advantage: time-to-market in weeks rather than months, latencies below 50 ms, no embedding service of your own, no vector store of your own. Limit: embedding models are vendor-proprietary, re-ranking logic is partly a black box, adaptation to industry-specific vocabulary happens via tuning rather than model swaps. For classic eCommerce search in B2B this is the pragmatic path in most cases.

Path B: Your own embedding pipeline on Snowflake/Cortex. Anyone already running on Snowflake — increasingly the case among foobar clients — can generate embeddings natively via Cortex (EMBED_TEXT_768, EMBED_TEXT_1024) and operate them with Cortex Search as a managed vector service. This opens use cases beyond the storefront: semantic search over datasheets, service manuals, contract documents, internal knowledge bases. Economically relevant: Cortex Search bills embedding indexing per token and index hosting per GB-month — which means it costs even without queries. That is a model you have to represent honestly in the calculation.

For agentic search the focus today is clearly on the second path: LLM agent (via OpenAI, Anthropic or Snowflake Cortex AI SQL) plus tool calls into the commerce platform. We go deeper on the data and AI pipeline in our parallel series on Snowflake — see the post Snowflake meets AI and LLMs. Further reading on agent patterns in commerce: foobar.agency/agentic-commerce.

The three limits most frequently underestimated in B2B implementations do not come from the technology but from the procurement context.

Hallucination. An LLM agent that recommends a product to a buyer that is not in the catalog is not producing a "creative suggestion" — it is producing a complaint case and, in the worst case, a compliance breach. Agentic search in B2B needs hard tool constraints: the agent may only recommend from the released catalog, every recommendation must be validated against an authoritative source of truth, the price comes from the pricing engine, not from the model.

Compliance. In industrial B2B, catalogs are shaped by regulation — REACH, RoHS, export control, industry-specific standards. An agent that recommends an item to a customer in country X that may not be shipped there is not "almost right". Compliance filters must take effect before the LLM response, not after.

Explainability. Buyers want to know why a specific product is recommended — to a supervisor, an auditor, the QA department. "The model produced it" does not carry. A productive B2B agent must be able to surface sources, datasheet references and the filters applied. Where that is not feasible, the classic result list with facets remains the cleaner approach.