The Retriever: Standardizing Search

In Module 6, we called db.similarity_search(). This is fine for scripts, but in an LCEL Chain (Module 4), we need a Retriever. A Retriever is a light wrapper around a vector store that follows the "Runnable" protocol.

1. Creating a Retriever

You don't need a new library. Every LangChain vector store has an .as_retriever() method.

# Convert your local DB into an AI-ready Retriever
retriever = db.as_retriever(search_kwargs={"k": 3})

# Now you can use '.invoke()' directly
docs = retriever.invoke("What is the refund policy?")

---

2. Why not just use db?

1. Uniformity: All retrievers worked the same way, whether they use a Vector DB, a SQL DB, or a Web Search API.
2. Chaining: You can pipe a retriever into a prompt using LCEL.
3. Advanced Logic: You can create "Multi-query" or "Self-query" retrievers that perform complex search math behind the scenes.

---

3. Visualizing the Retriever Node

graph LR
    Input[User Query String] --> Ret[Retriever Object]
    Ret -->|Search| V[Vector Store]
    V -->|List| Docs[List of Document Objects]

---

4. Engineering Tip: Multi-Query Retrieval

Sometimes the user asks a "Bad" question.

• User: "Policy?" (Too vague).
• Multi-Query Retriever: Generates 3 variations ("What is the refund policy?", "Our company policy on returns", etc.), searches the DB for all three, and combines the results. This significantly improves accuracy.

---

5. Metadata Filtering

You can configure your retriever to only look at specific types of data.

retriever = db.as_retriever(
    search_kwargs={'filter': {'category': 'HR'}}
)

---

Key Takeaways

• .as_retriever() turns a database into a chainable search node.
• The output of a Retriever is always a List[Document].
• Search Kwargs allow you to set k and filters at the node level.
• Retrievers are the "Connectors" between your data and your chains.