Skip over navigation

Embeddings

Numerical vector representations of text, images, or other data that capture semantic meaning, allowing a machine to measure how similar two pieces of content are by comparing their vectors.

Created Sep 10, 2026
Data and observability

What It Is

An embedding is produced by a model that converts input into a fixed-length list of numbers positioned in a high-dimensional space, such that similar concepts land near each other in that space.

Embeddings are the foundation underneath vector search, AI-driven recommendations, and retrieval for RAG, any feature that depends on measuring ‘how similar’ two things are.

They are not human-readable. A vector is just numbers; similarity between two embeddings is computed mathematically, not inspected directly.

In B2B Commerce Context

In B2B commerce, the same product is frequently described in incompatible ways by different suppliers, and embeddings are what lets a system treat those descriptions as the same underlying concept.

A practical example

Two SKUs described completely differently by two suppliers, '3/4 in. copper elbow fitting' versus 'copper pipe bend 90°, 0.75in', produces embeddings that sit close together, letting AI catalog management or search treat them as effectively equivalent without manual mapping.

This underlies semantic search, product deduplication, and the retrieval step in any RAG-based support or merchandising feature.

When You Need It

  • You're building semantic search or product-matching features.
  • You need to deduplicate catalog entries with inconsistent naming across suppliers.
  • You're powering AI agent retrieval (RAG) over product or policy data.
  • Any feature you're building relies on 'similar to' logic rather than exact matching.

Be consistent about which model generates your embeddings, mixing sources in one index degrades results.

What It Is Not

  • Embeddings are not human-readable or directly interpretable. Similarity is computed, not inspected.
  • An embedding is not a database itself. Embeddings are the data; a vector database is where they're stored and searched.
  • They are not permanent or universal. Embeddings depend on the model that generated them, and mixing different models' embeddings in one index produces inconsistent results.

Comparison

Capability Embeddings-Based Matching Keyword Matching
Basis for match Semantic meaning captured as vectors Exact or partial text match
Handles synonyms/variation Yes, inherently Only with a manually maintained list
Interpretability Not directly readable by humans Directly readable
Consistency requirement Must use a consistent embedding model No equivalent constraint

See also

Ready to see it in action?

Book a demo of OroCommerce

See how agentic workflows fit into complex B2B commerce, with a walkthrough tailored to your stack.

Book a demo

Share

Back to top