A model only knows what it was trained on. To answer from your own notes, a system first retrieves the passages that fit the question, then gives them to the model with the question. This is retrieval-augmented generation, RAG.

Real systems use learned embeddings. The idea works with the simplest vectors too: count each word.

import re

text = "Cats sleep. Cats purr."
words = re.findall(r"[a-z]+", text.lower())
print(words)
# ['cats', 'sleep', 'cats', 'purr']

Give every word in the documents a position, turn each text into a vector of counts, and rank the documents by cosine similarity to the question. The best match goes into the prompt.

Write words(text) returning the lowercase words (runs of letters a to z). Write vectorize(text, vocab) returning a float tensor with the count of each vocabulary word in the text, in vocab's order. Write retrieve(question, docs, k=1) that builds the vocabulary from all the documents' words (sorted), and returns the k documents most similar to the question by cosine similarity, best first.

The explanation opens here once every check passes.