Persisting the Vector Store

Where we left off

Every lesson so far rebuilds the vector store from scratch, embedding every chunk again, every single time the script runs. That's wasteful (each run burns embedding calls against a rate-limited free tier for work that produced the exact same result last time) and it means an in-memory Python list, true to its name, vanishes the moment the program exits. This lesson fixes both with the simplest possible persistence: a JSON file.

Why JSON is enough here

A vector, underneath the abstraction, is just a list of floating-point numbers. A vector store, in this course, is just a list of dicts. JSON already knows how to represent exactly those two things, lists and dicts of strings/numbers, with no translation step needed:

path.write_text(json.dumps(store))

This won't scale to millions of records (loading the whole file back into memory becomes slow and memory-hungry well before that point), but for the handful of documents this course works with, it's a complete, honest answer to "how do I not lose my embeddings when the program exits."

The code, piece by piece

def save_store(store: list[dict], path: Path) -> None:
    path.write_text(json.dumps(store))

def load_store(path: Path) -> list[dict]:
    return json.loads(path.read_text())

Two small, symmetric functions. save_store turns the Python list into a JSON string and writes it to disk; load_store reads that string back and turns it back into the exact same Python list, floats and all.

if STORE_PATH.exists():
    store = load_store(STORE_PATH)
    ...
else:
    store = build_vector_store()
    save_store(store, STORE_PATH)

The pattern every persistence layer follows in some form: check if saved data exists first, load it if so; only do the expensive work (embedding, here) if it doesn't.