A real model and a real embedder

Eighteen lessons, no key. Here is what changes when you use one, and it is two lines in one file.

The stand-ins went in through the langchain provider, and so does anything else. Replace the two instances and nothing else in the course moves.

from langchain_openai import ChatOpenAI, OpenAIEmbeddings

config = {
    "llm": {"provider": "langchain", "config": {"model": ChatOpenAI(model="gpt-5-mini")}},
    "embedder": {"provider": "langchain", "config": {"model": OpenAIEmbeddings()}},
    "vector_store": {"provider": "qdrant", "config": {
        "path": "memories", "embedding_model_dims": 1536}},
}

The dimension

embedding_model_dims changes with the embedder, and getting it wrong is the first error most people hit. The stand-in made 64 numbers; OpenAI's small embedding model makes 1536. Say 64 with a real embedder and the store rejects the first insert.

A store that already holds memories at one length cannot take another. Changing embedder means re-embedding everything, which in practice means adding it all again into a new directory.

What else changes

Extraction gets much better. The stand-in keeps sentences; a real model rewrites them into standalone facts, merges duplicates, and notices when a new sentence contradicts an old one, which turns the UPDATE events from lesson 15 from theory into something you see.

Search starts matching meaning. The limitation from lesson 6 disappears: how should we contact him finds the email memory without sharing a word with it. This is the single biggest visible difference.

Custom instructions start working. Lesson 17 proved the rule reaches the model; with a real one behind it, the rule is obeyed.

It costs money and time. Every add is a model call and every memory is an embedding call. A chat with twenty turns is twenty extractions unless you batch them, which is the main thing to think about before putting this in a loop.

Keep the stand-in

Do not delete it. A local model and embedder make a test suite that runs on every commit for nothing, which is exactly what lesson 22 needs.