RAG: Embeddings and Vectorstore

Where we left off

Lesson 27 loaded notes.txt and split it into six focused chunks. This lesson makes those chunks searchable, not by exact keyword matching, but by meaning.

Embeddings: turning text into numbers that capture meaning

embeddings_model = GoogleGenerativeAIEmbeddings(model="models/gemini-embedding-001")

An embedding is a long list of numbers (a "vector") that represents a piece of text. The key property that makes this useful: texts with similar meaning end up with similar numbers, positioned close together in that number-space, even if they don't share a single exact word.

"A dog ran in the park" and "A puppy played outside" would embed to nearby vectors, despite having almost no words in common, because their meanings are close. This is a genuinely different idea from anything earlier in this course, everything before this compared or searched text, this compares meaning.

A vector store: a searchable collection of embedded chunks

vector_store = InMemoryVectorStore(embeddings_model)
vector_store.add_documents(chunks)

A vector store holds a collection of chunks, each one embedded into its vector form, and knows how to efficiently find the chunks whose vectors are closest to a new query's vector. InMemoryVectorStore keeps everything in your program's memory, no separate database needed, which makes it perfect for learning and small projects, though a production system with millions of documents would use a dedicated vector database instead.

Searching by meaning, not by words

query = "What do I know about baking bread at home?"
results = vector_store.similarity_search(query, k=2)

Here's the part worth sitting with. The query asks about "baking bread," but nothing in notes.txt mentions bread at all, the closest matching chunk is actually about pizza dough. A plain keyword search (like searching a text file with Ctrl+F) would find nothing, "bread" doesn't appear anywhere in the file.

But run this lesson, and the pizza dough chunk comes back as the top match anyway, because baking pizza dough and baking bread are conceptually close: both involve dough, yeast, fermentation, and oven temperature. The embedding captured that closeness even though the exact words don't overlap.

k=2 asks for the two closest chunks, not just one, in case more than one chunk is relevant to a question.

Why this matters for RAG

This is the actual mechanism that makes RAG work: instead of stuffing an entire document into the model's context every time (expensive, and eventually impossible past the context window, recall Lesson 25), you embed the question, find just the few chunks that are actually relevant to it, and only send those specific chunks to the model. Lesson 29 does exactly that, wrapping this search as a tool an agent can use.