Understanding Embeddings

What Are Embeddings?

Embeddings are dense vector representations of text (or other data) in a high-dimensional space where semantic similarity corresponds to geometric proximity. They are the foundation of RAG systems.

How Embeddings Work

• Text is passed through an embedding model (e.g., OpenAI's text-embedding-3-small, Cohere embed, BGE)
• The model outputs a fixed-size vector (e.g., 1536 dimensions)
• Similar texts produce vectors that are close together in this space
• Distance metrics (cosine similarity, dot product) measure semantic relatedness

Example

# These would have high cosine similarity:
embed("How do I reset my password?")  ≈  embed("I forgot my login credentials")

# These would have low cosine similarity:
embed("How do I reset my password?")  ≠  embed("What's the weather today?")

Choosing an Embedding Model

Key Considerations

• Embedding models are different from generation models — they're optimized for representation, not text output
• Once you choose an embedding model, switching requires re-embedding all your data
• Dimension reduction (e.g., Matryoshka embeddings) can reduce storage without much quality loss

🌼 Daisy+ in Action: Semantic Understanding

Daisy+ uses embeddings to power smart search across the ERP — product catalogs, knowledge base articles, and historical customer interactions are all embedded and indexed, enabling semantic search that understands intent rather than just keywords. A customer asking about "affordable project tracking" can find relevant products even if those exact words don't appear in the product description.