Database indexing and vector databases · Chapter 8

Why indexes make search fast, B-tree vs vector indexes, and what a vector database like Pinecone, pgvector or Neo4j does.

The big idea

An index lets a database find things without reading every row, just like the index at the back of a book lets you jump to the right page. With 50 nodes it does not matter. With 5 million, it is the difference between a millisecond and a minute.

Two kinds of index

Kind	Answers questions like	How it works, simply
Regular index (B-tree)	"Find the user with this email"	Keeps values sorted, like a phone book, so an exact match is found in a few steps.
Vector index	"Find the 5 texts closest in meaning to this one"	Groups vectors that are near each other, so a search only checks a small neighbourhood instead of everything.

A vector index gives an approximate answer: it may very rarely miss a close match, in exchange for being hugely faster. For almost every app that trade is worth it.

What is a vector database?

A database that stores embeddings and has a vector index built in. Popular choices:

Dedicated vector databases: Pinecone, Qdrant, Weaviate, Chroma.
Normal databases that added vector search: PostgreSQL with pgvector, and Neo4j.

GraphRAG Chats uses Neo4j because it stores the graph and the vectors in one place, so one database can do both the meaning search and the relationship walk.

Where it lives in GraphRAG Chats

When the backend starts, ensure_vector_index in backend/src/backend/neo4j_client.py creates the index:

cypher

CREATE VECTOR INDEX node_embedding_index IF NOT EXISTS
FOR (n:GraphNode) ON (n.embedding)
OPTIONS {indexConfig: {
  `vector.dimensions`: 1024,
  `vector.similarity_function`: 'cosine'
}}

In plain words: index the embedding of every GraphNode, expect 1024 numbers, and compare them with cosine similarity (Chapter 7).

Practice: find the index

Open the Neo4j Browser at http://localhost:7474.
Run SHOW INDEXES and find node_embedding_index. Check that its type is VECTOR.
Find the search query in neo4j_client.py: it starts with CALL db.index.vector.queryNodes. That one line is the vector search.

You are done when you can point to the index in Neo4j and to the line of code that uses it.

Check yourself

Why does a database need an index?
What is the trade-off a vector index makes?
Why does GraphRAG Chats use Neo4j instead of a separate vector database?