Retrieval: finding the right documents

Retrieval is finding the few documents that answer a question, so the model can read them instead of guessing. Here you build a tiny keyword retriever to make the idea concrete.

Last updated: 27 Sep, 2026 · LangGraph 1.2

A model does not know your policies or your data. Retrieval fetches the relevant lines first; the next lesson has the model answer from them.

The documents

Start with the documents to search. Here they are three short support lines.

docs = [
    "Refunds are processed within 5 working days.",
    "Reset your password from the settings page.",
    "Orders ship within 2 working days.",
]

The stop words

Some words carry no meaning for matching, such as "how" and "the". Put those in a stop set so the search ignores them.

STOP = {"how", "long", "for", "a", "the", "to", "do", "you", "is"}   # words to ignore

The search function

The search splits the question into words, drops the stop words, and keeps any document that contains one of the remaining words.

def search(query, docs):
    terms = [w for w in query.lower().split() if w not in STOP]   # content words only
    return [d for d in docs if any(t in d.lower() for t in terms)]  # docs that share one

Asking a question

Ask a question. After the stop words are dropped, "refund" is the word that matters, and it appears in the first document.

print(search("how long for a refund", docs))

The retriever in a run

The same pieces in one file, ready to run.

docs = [
    "Refunds are processed within 5 working days.",
    "Reset your password from the settings page.",
    "Orders ship within 2 working days.",
]

STOP = {"how", "long", "for", "a", "the", "to", "do", "you", "is"}

def search(query, docs):
    terms = [w for w in query.lower().split() if w not in STOP]
    return [d for d in docs if any(t in d.lower() for t in terms)]

print(search("how long for a refund", docs))

What the search returned

• The retriever keeps any document that shares a content word with the question; common words like "how" and "for" are ignored.
• "refund" matched the first line, so only that one came back.
• A real retriever uses embeddings to match by meaning, not exact words, but the shape is the same: question in, a short list of documents out.

Keyword vs embeddings

When to retrieve documents

• Answering from a company's own documents, policies or a knowledge base.
• The retrieve step of a RAG pipeline, before the model reads the results.

Related

• Previous: Capstone: the support agent
• Next: RAG: answering from what you found