Hour 5 - Working with Vector Databases

Lecture Notes: 

---

1. Concepts

What is a Vector Database?

• A vector database is designed to store, manage, and query high-dimensional vector embeddings efficiently.
• It enables similarity search and nearest neighbor queries, critical for working with embeddings generated by models like Ollama.

Key Features of Vector Databases:

1. High-Dimensional Indexing: Stores embeddings (vectors) and enables fast searches.
2. Similarity Search: Finds vectors closest to a given query vector using distance metrics like cosine similarity or Euclidean distance.
3. Scalability: Handles large-scale data efficiently.
4. Integration: Can work with other data structures (e.g., JSON metadata) for richer querying.

---

2. Key Aspects

Why Use a Vector Database?

• Efficiency: Optimized for querying large-scale vector data.
• Accuracy: Provides precise similarity results using advanced indexing algorithms like HNSW (Hierarchical Navigable Small World).
• Real-World Use Cases: Image retrieval, semantic search, recommendation systems, chatbots.

Common Vector Databases:

• Pinecone
• Weaviate
• Qdrant
• Milvus

Querying Techniques:

1. K-Nearest Neighbors (KNN): Finds top K vectors closest to the query vector.
2. Hybrid Search: Combines vector similarity with traditional keyword-based searches.

---

3. Implementation

Setting Up a Vector Database

1. Install and configure the database. Most vector databases provide cloud-hosted and local setups.
2. Store embeddings: Use the embeddings generated by ollama embed.
3. Query embeddings: Perform similarity searches to find relevant results.

---

4. CLI Commands for Working with Vector Databases

Command	Description	Example
ollama embed	Generate embeddings for text or documents.	ollama embed "example text" --format json
ollama run	Use embeddings as part of a model query.	ollama run mymodel --format json

---

5. Real-Life Example

Scenario: Build a Semantic Search Engine with a Vector Database

Suppose we want to search through a collection of customer reviews to find those most relevant to a user query. The embeddings of the reviews will be stored in a vector database and queried for similarity.

---

6. Code Examples

Step 1: Install Qdrant Vector Database

Qdrant is an easy-to-use vector database with local and cloud options.

# Install Qdrant locally via Docker
docker pull qdrant/qdrant
docker run -p 6333:6333 qdrant/qdrant

Step 2: Store Embeddings in Qdrant

import qdrant_client
from qdrant_client.models import PointStruct
import json

# Initialize Qdrant client
client = qdrant_client.QdrantClient(url="http://localhost:6333")

# Create a collection for embeddings
client.recreate_collection(
    collection_name="customer_reviews",
    vector_size=512,  # Dimension of embeddings
    distance="Cosine"
)

# Load embedding generated by Ollama
with open("review1_embedding.json", "r") as file:
    review1 = json.load(file)

with open("review2_embedding.json", "r") as file:
    review2 = json.load(file)

# Insert embeddings into the database
points = [
    PointStruct(id=1, vector=review1["embedding"], payload={"text": review1["text"]}),
    PointStruct(id=2, vector=review2["embedding"], payload={"text": review2["text"]}),
]

client.upsert(collection_name="customer_reviews", points=points)

Step 3: Query the Vector Database

# Simulate a user query
query = "What do customers say about product quality?"

# Generate query embedding (replace with actual embedding generated by Ollama)
query_embedding = [0.12, 0.34, ...]  # Placeholder example

# Perform similarity search
results = client.search(
    collection_name="customer_reviews",
    query_vector=query_embedding,
    limit=2  # Retrieve top 2 matches
)

# Display results
for result in results:
    print(f"Score: {result.score}")
    print(f"Review: {result.payload['text']}")

---

7. Summary

• Concepts Covered: What vector databases are, why they're useful, and how they enable efficient similarity searches.
• Key Aspects: High-dimensional indexing, similarity measures, and practical use cases.
• CLI Commands: Use ollama embed to generate embeddings for storing in the database.
• Real-Life Example: Semantic search through customer reviews using Qdrant.
• Code Examples: Storing and querying embeddings in Qdrant.

---

8. Homework/Practice

1. Install a vector database of your choice (e.g., Qdrant, Milvus).
2. Generate embeddings for five text samples using ollama embed.
3. Store these embeddings in the database.
4. Implement a Python script to perform similarity searches and rank the results.
5. Experiment with different distance metrics (e.g., Euclidean vs. Cosine).

---

This lecture introduces students to working with vector databases and includes a practical example with Qdrant, a widely-used vector database.