Searching Vector Stores

🔍 Vector Search and RAG: Finding Your Perfect Match

The Library Analogy

Imagine you have a magical library with millions of books. But this library is special. Instead of organizing books by title or author, it organizes them by what they mean.

When you ask a question, the librarian doesn’t just look for your exact words. She finds books that understand your question — even if they use different words!

This is Vector Search. And when we combine it with AI to generate answers? That’s RAG (Retrieval-Augmented Generation).

🎯 What is a Vector Store?

Think of it like a memory palace for your AI.

Every piece of text (document, sentence, paragraph) gets converted into a list of numbers called a vector. These numbers capture the meaning of the text.

# A sentence becomes numbers!
"I love pizza" → [0.2, 0.8, 0.1, ...]
"Pizza is delicious" → [0.3, 0.7, 0.2, ...]
# Similar meanings = similar numbers!

Why does this matter?

• Similar meanings = vectors that are close together
• Different meanings = vectors that are far apart

🔎 Similarity Search

The Story

You’re in that magical library. You whisper: “Tell me about healthy breakfast ideas.”

The librarian runs through the library and brings back:

1. “Oatmeal recipes for morning energy”
2. “Nutritious ways to start your day”
3. “Breakfast smoothie guide”

She didn’t search for your exact words. She found books with similar meaning.

How It Works

# Your question becomes a vector
query = "healthy breakfast ideas"

# Find the 3 most similar documents
results = vectorstore.similarity_search(
    query,
    k=3  # Return top 3 matches
)

The Magic Behind It

graph TD
    A["Your Question"] --> B["Convert to Vector"]
    B --> C["Compare with All Stored Vectors"]
    C --> D["Find Closest Matches"]
    D --> E["Return Top Results"]

Key Point: The closer two vectors are, the more similar their meaning!

🎪 Maximum Marginal Relevance (MMR)

The Problem

Imagine asking for “pizza recipes” and getting:

1. Pizza dough recipe
2. Pizza dough recipe (slightly different)
3. Pizza dough recipe (from another book)

All relevant, but boring and repetitive!

The Solution: MMR

MMR is like a smart librarian who says:

“I’ll give you relevant results, but I’ll also make sure they’re different from each other!”

# Get diverse results, not just similar ones
results = vectorstore.max_marginal_relevance_search(
    query="pizza recipes",
    k=3,           # Return 3 results
    fetch_k=10,    # Consider top 10 first
    lambda_mult=0.5  # Balance: relevance vs diversity
)

What You Get Instead

1. Pizza dough recipe
2. Pizza sauce techniques
3. Pizza topping combinations

Same topic, different angles!

Understanding lambda_mult

🏷️ Metadata Filtering

The Story

Back to our magical library. You say:

“I want pizza recipes, but ONLY from Italian cookbooks published after 2020.”

The librarian now has two jobs:

1. Find books about pizza (similarity search)
2. Filter by your rules (metadata filtering)

What is Metadata?

Extra information attached to each document:

# A document with metadata
{
    "content": "How to make perfect pizza dough",
    "metadata": {
        "author": "Mario Rossi",
        "year": 2022,
        "cuisine": "Italian",
        "difficulty": "easy"
    }
}

Using Filters

# Search with filters
results = vectorstore.similarity_search(
    query="pizza recipes",
    k=3,
    filter={
        "cuisine": "Italian",
        "year": {"$gte": 2020}  # >= 2020
    }
)

Common Filter Operations

🔀 Hybrid Search

The Best of Both Worlds

Remember our two search methods?

1. Keyword search (traditional): Finds exact words
2. Vector search (semantic): Finds similar meanings

Hybrid search combines both!

Why Do We Need Both?

How It Works

graph TD
    A["Your Query"] --> B["Keyword Search"]
    A --> C["Vector Search"]
    B --> D["Merge Results"]
    C --> D
    D --> E["Rank & Return Best"]

Code Example

# Many vector stores support hybrid search
results = vectorstore.hybrid_search(
    query="Python TypeError fix",
    k=5,
    alpha=0.5  # 0 = keyword only, 1 = vector only
)

When to Use What?

🔄 Vector Store as Retriever

The Big Picture

In LangChain, a Retriever is like a helper that fetches documents for your AI. And guess what? Your vector store can become one!

Why Make It a Retriever?

Think of it like this:

• Vector Store = The library
• Retriever = The librarian who knows how to use the library

When you turn a vector store into a retriever, you can:

• Plug it into LangChain chains
• Use it in RAG pipelines
• Combine it with other tools

Simple Conversion

# Turn your vector store into a retriever
retriever = vectorstore.as_retriever()

# Now use it!
docs = retriever.invoke("What is pizza?")

Customizing Your Retriever

# With search parameters
retriever = vectorstore.as_retriever(
    search_type="similarity",  # or "mmr"
    search_kwargs={
        "k": 5,
        "filter": {"category": "recipes"}
    }
)

Different Search Types

Using in a RAG Chain

from langchain.chains import RetrievalQA

# Create a Q&A chain with your retriever
qa_chain = RetrievalQA.from_chain_type(
    llm=my_llm,
    retriever=retriever,
    chain_type="stuff"
)

# Ask questions!
answer = qa_chain.invoke(
    "What's the best pizza dough recipe?"
)

graph TD
    A["User Question"] --> B["Retriever"]
    B --> C["Fetch Relevant Docs"]
    C --> D["Pass to LLM"]
    D --> E["Generate Answer"]
    E --> F["Return to User"]

🎯 Putting It All Together

Here’s how all these pieces work in a real RAG system:

from langchain_community.vectorstores import Chroma
from langchain_openai import OpenAIEmbeddings

# 1. Create vector store
vectorstore = Chroma.from_documents(
    documents=my_docs,
    embedding=OpenAIEmbeddings()
)

# 2. Convert to retriever with MMR
retriever = vectorstore.as_retriever(
    search_type="mmr",
    search_kwargs={
        "k": 4,
        "fetch_k": 10,
        "filter": {"verified": True}
    }
)

# 3. Use in RAG pipeline
relevant_docs = retriever.invoke(
    "How do I make pizza dough?"
)

📝 Quick Summary

🚀 You’ve Got This!

Vector search isn’t magic — it’s just smart math that understands meaning. Now you know:

• ✅ How to find similar documents
• ✅ How to get diverse results with MMR
• ✅ How to filter by metadata
• ✅ How to combine keyword and semantic search
• ✅ How to use vector stores in LangChain

Go build something amazing! 🎉