Embedding

An embedding model takes an input, typically a sentence, paragraph or document chunk, and outputs a fixed-length vector, often several hundred to a few thousand numbers. The model is trained so that inputs with similar meaning produce vectors that are close in that space. "Refund not received" and "money hasn't come back to my account" land near each other even though they share no words.

Search works by embedding the query the same way and finding the stored vectors nearest to it, usually by cosine similarity. That is what powers semantic search and the retrieval half of RAG. Embeddings are stored and indexed in a vector database or an extension such as pgvector.

Embeddings are not understanding and not keyword matching. They compress meaning lossily, so exact identifiers (invoice numbers, SKUs, error codes) are often matched badly, which is why hybrid search pairs them with a keyword index. Embeddings from different models are not comparable; changing the embedding model means re-indexing everything.

• CTO / Head of Engineering: The choice of embedding model fixes the retrieval ceiling and locks the index; switching later is a full re-index, so choose with a benchmark on your own data.
• Data lead: Embeddings turn unstructured text into something you can cluster, deduplicate and search alongside structured fields.
• CISO: Embeddings sent to a hosted API are derived from your documents; for private data the model may need to run inside your perimeter.
• CFO: Embedding cost is a one-off per document plus a small cost per query, not a recurring licence, but re-indexing large corpora is not free.

Computers compare strings, not meanings. Keyword search misses a question phrased differently from the document, and rules-based synonym lists never keep up. Embeddings exist to give software a usable notion of "means roughly the same thing" so retrieval, deduplication and recommendation can work on intent. The trade-off is imprecision: embeddings blur detail, struggle with numbers and codes, and can rank a plausible-sounding but wrong passage above the exact one. They are a component of a retrieval system, not the whole of it.

• Matching support tickets to help-centre articles in a SaaS product regardless of how the customer phrased the problem
• Grouping near-duplicate KYC documents and detecting resubmissions at an NBFC
• Catalogue search for a retailer where "warm jacket for Bengaluru winter" returns light fleece rather than heavy parkas
• Finding similar past cases in a hospital's protocol library for a clinician's query
• Clustering learner questions in an education platform to find gaps in course material

ConceptA foundational concept in retrieval and personalisation work. Selecting, benchmarking and hosting embedding models is part of Eazyware's Retrieval & Knowledge Engineering and AI/ML Development services.

Eazyware service that covers it: Retrieval & Knowledge Engineering. Starting prices are on the pricing page.