Learn
What is an embedding?
An embedding is a dense vector representation of text, image, or other data — numbers that capture semantic meaning so similar concepts sit close together in vector space. Embeddings power semantic search, RAG retrieval, and clustering in agent memory.
In depth
Language models read tokens; databases and search indexes need comparable geometry. An embedding model maps a chunk of text to a fixed-length vector (often 768–3072 dimensions). Chunks with similar meaning produce vectors with high cosine similarity.
In agent systems, embeddings typically appear in three places: retrieval — find relevant docs for RAG before the model answers; memory — store past decisions and fetch what matters for the current task; routing — classify intent or pick which skill/role should handle a request.
Quality depends on the embedding model, chunking strategy, and metadata filters. Small chunks improve precision; larger chunks improve context. Hybrid search (embeddings + keyword) often beats either alone.
Neuro OS roles combine layered memory with connector-backed retrieval — embeddings are an implementation detail behind scoped knowledge, not something every operator tunes by hand.
Examples
- Support KB — embed ticket history; retrieve similar resolved cases
- Legal review — embed contract clauses; surface precedent paragraphs
- Sales — embed prospect notes; match to winning deal patterns
- OpenAI text-embedding-3 — common API for production RAG
- Chunk + metadata filter — 'only finance docs from 2025'
Related terms
FAQ
Are embeddings the same as fine-tuning?
No. Fine-tuning changes model weights. Embeddings are separate vectors used for search and retrieval without retraining the LLM.
How big should chunks be?
Typical range 256–1024 tokens. Tune on your eval set — legal text wants different chunking than chat logs.
Run governed agent roles on a company OS — not only definitions in a glossary.