An embedding represents text as a vector: an ordered sequence of numbers that software can compare. Words, phrases, or documents with related meanings can have nearby representations in a vector space. This makes embeddings useful when an enterprise application needs to find relevant information by meaning rather than merely reuse the wording of a query. section 2.8.4 section 10.4.2
From text to numbers
Illustrative vectors might represent “cat” as [0.2, 0.3, -0.1] and “dog” as [0.4, 0.1, 0.2]. Smaller distances suggest stronger relatedness; a similarity score makes that relationship usable in search. Embeddings are learned from text during model training, so these numbers are representations produced by a model, not labels assigned by hand. section 2.8.4
Choose an embedding approach for the task: similarity embeddings compare pieces of text, text-search embeddings relate a short query to longer documents, and code-search embeddings relate code snippets to natural-language queries. Keep vectors from one provider’s embedding method with that method; another provider’s method cannot interpret them interchangeably. section 2.8.4
Listing 2.3 Getting an embedding in OpenAI
import os
from openai import OpenAI
client = OpenAI(api_key=’your-API-key’)
def get_embedding(text):
response = client.embeddings.create(
model="text-embedding-ada-002",
input=text)
return response.data[0].embedding
embeddings = get_embedding("I have a white dog named Champ.")
print("Embedding Length:", len(embeddings))
print("Embedding:", embeddings[:5])
A short Python listing calls OpenAI’s embeddings API with text-embedding-ada-002, then prints the returned vector’s length and first five values. Its hard-coded API-key placeholder is illustrative, not a production credential-handling pattern. section 2.8.4
For retrieval-augmented generation (RAG), process source documents into chunks, create their embeddings, and save them in a vector database. When a user asks a question, embed the question, search for similar stored chunks, and add the retrieved text to the LLM prompt. A chat application over years of blog posts uses Redis and a hybrid search index to retrieve matching chunks before generating an answer. section 8.1.2 section 8.6
Avoid recreating embeddings for every question when the source can be processed and indexed ahead of time: repeated work adds latency and cost. In production, plan chunking around the data, and use a vector index and database to manage performance, scale, and security. Building the full retrieval-and-chat workflow still requires substantial data engineering and integration work. section 8.1.2 section 8.9
