Embeddings & Semantic Search in Python (OpenAI) | AI Engineer Course #8

0:00 Keyword search fails the moment people use different words. Search for how do I get my money back and a page titled refund policy might never appear. Semantic search fixes that by searching by meaning. In this lesson, you will create embeddings, compare them with a little maths, and build a working semantic search engine in about 40 lines of Python. In lesson one, we saw that models turn tokens into embeddings, lists of numbers that capture meaning. OpenAI lets you create embeddings for whole sentences or paragraphs. Texts with similar meanings get vectors that point in similar directions, even if they share no words at all. The model we will use, text-embedding-3-small, turns any text into a list of 1,536 numbers.

0:51 Creating an embedding is one call, client.embeddings.create, with the embedding model and your text. The vector is in response.data.embedding. Print its length and you get 1,536. The numbers themselves mean nothing to a human. What matters is how they compare with other vectors. You can pass a whole list of texts in one request, which is much faster than one call per text. Here are four short help center documents. We embed them all at once and collect the vectors with a list comprehension. In a real app, you would do this once, save the vectors, and reuse them, rather than paying to embed the same documents again every time.

1:38 How do we compare two vectors? With cosine similarity, which measures the angle between them. Close to one means they point the same way, so the meanings are similar. Close to zero means unrelated. With NumPy, it is one line, the dot product divided by the lengths of the two vectors. Openai embeddings are already normalized to length one, so in practice the dot product alone gives the same answer. Now the search engine. Embed the user's query, score it against every document vector, sort from highest to lowest, and return the top results. Search for how do I get my money back, and the refund document comes out on top, even though it never uses the words money or back.

2:23 That is semantic search, matching by meaning. Embedding documents costs time and money, so save the results. For a small project, a JSON file is perfectly fine. Store the texts and their vectors, and load them when your app starts. When you have many thousands of documents, you move to a vector database such as Chroma, PGVector, or Pinecone, which store vectors and search them very fast. The idea is exactly the same as our 40 lines of code.

2:54 Embeddings power far more than search. They are the retrieval step in rag, which we build next lesson. They also drive recommendations, more like this features, finding duplicate tickets, and grouping feedback into themes. Once you can turn text into vectors and compare them, a whole category of AI features opens up. You have built a semantic search engine from scratch. Next lesson, we connect it to a language model and build a full rag app that answers questions from your own documents with sources. Don't miss it.