yukaty

Posted on Nov 12

Part 3: Implementing Vector Search with Ollama

#ai #postgres #docker #opensource

Part 1 covered PostgreSQL with pgvector setup, and Part 2 implemented vector search using OpenAI embeddings. This final part demonstrates how to run vector search locally using Ollama! ✨

Contents
Why Ollama?
Setting Up Ollama with Docker
Database Updates
Implementation
Search Queries
Performance Tips
Troubleshooting
OpenAI vs. Ollama
Wrap Up

Why Ollama? 🦙

Ollama allows you to run AI models locally with:

Offline operation for better data privacy
No API costs
Fast response times

We'll use the nomic-embed-text model in Ollama, which creates 768-dimensional vectors (compared to OpenAI's 1536 dimensions).

Setting Up Ollama with Docker 🐳

To add Ollama to your Docker setup, add this service to compose.yml:

services:
  db:
    # ... (existing db service)

  ollama:
    image: ollama/ollama
    container_name: ollama-service
    ports:
      - "11434:11434"
    volumes:
      - ollama_data:/root/.ollama

  data_loader:
    # ... (existing data_loader service)
    environment:
      - OLLAMA_HOST=ollama
    depends_on:
      - db
      - ollama

volumes:
  pgdata:
  ollama_data:

Then, start the services and pull the model:

docker compose up -d

# Pull the embedding model
docker compose exec ollama ollama pull nomic-embed-text

# Test embedding generation
curl http://localhost:11434/api/embed -d '{
  "model": "nomic-embed-text",
  "input": "Hello World"
}'

Database Updates 🐘

Update the database to store Ollama embeddings:

-- Connect to the database
docker compose exec db psql -U postgres -d example_db

-- Add a column for Ollama embeddings
ALTER TABLE items
ADD COLUMN embedding_ollama vector(768);

For fresh installations, update postgres/schema.sql:

CREATE TABLE items (
    id SERIAL PRIMARY KEY,
    name VARCHAR(255) NOT NULL,
    item_data JSONB,
    embedding vector(1536),        # OpenAI
    embedding_ollama vector(768)   # Ollama
);

Implementation 👾

Update requirements.txt to install the Ollama Python library:

ollama==0.3.3

Here’s an example update for load_data.py to add Ollama embeddings:

import ollama  # New import

def get_embedding_ollama(text: str):
    """Generate embedding using Ollama API"""
    response = ollama.embed(
        model='nomic-embed-text',
        input=text
    )
    return response["embeddings"][0]

def load_books_to_db():
    """Load books with embeddings into PostgreSQL"""
    books = fetch_books()

    for book in books:
        description = (
            f"Book titled '{book['title']}' by {', '.join(book['authors'])}. "
            f"Published in {book['first_publish_year']}. "
            f"This is a book about {book['subject']}."
        )

        # Generate embeddings with both OpenAI and Ollama
        embedding = get_embedding(description)                # OpenAI
        embedding_ollama = get_embedding_ollama(description)  # Ollama

        # Store in the database
        store_book(book["title"], json.dumps(book), embedding, embedding_ollama)

Note that this is a simplified version for clarity. Full source code is here.

As you can see, the Ollama API structure is similar to OpenAI’s!

Search Queries 🔍

Search query to retrieve similar items using Ollama embeddings:

-- View first 5 dimensions of an embedding
SELECT
    name,
    (replace(replace(embedding_ollama::text, '[', '{'), ']', '}')::float[])[1:5] as first_dimensions
FROM items;

-- Search for books about web development:
WITH web_book AS (
    SELECT embedding_ollama FROM items WHERE name LIKE '%Web%' LIMIT 1
)
SELECT
    item_data->>'title' as title,
    item_data->>'authors' as authors,
    embedding_ollama <=> (SELECT embedding_ollama FROM web_book) as similarity
FROM items
ORDER BY similarity
LIMIT 3;

Performance Tips 📊

Add an Index

CREATE INDEX ON items
USING ivfflat (embedding_ollama vector_cosine_ops)
WITH (lists = 100);

Resource Requirements

RAM: ~2GB for the model
First query: Expect slight delay for model loading
Subsequent queries: ~50ms response time

GPU Support

If processing large datasets, GPU support can greatly speed up embedding generation. For details, refer to the Ollama Docker image.

Troubleshooting 🔧

Connection Refused Error

The Ollama library needs to know where to find the Ollama service. Set the OLLAMA_HOST environment variable in data_loader service:

data_loader:
  environment:
    - OLLAMA_HOST=ollama

Model Not Found Error

Pull the model manually:

docker compose exec ollama ollama pull nomic-embed-text

Alternatively, you can add a script to automatically pull the model within your Python code using the ollama.pull() function.

High Memory Usage

Restart Ollama service
Consider using a smaller model

OpenAI vs. Ollama ⚖️

Feature	OpenAI	Ollama
Vector Dimensions	1536	768
Privacy	Requires API calls	Fully local
Cost	Pay per API call	Free
Speed	Network dependent	~50ms/query
Setup	API key needed	Docker only

Wrap Up 🌯

This tutorial covered only how to set up a local vector search with Ollama. Real-world applications often include additional features like:

Query optimization and preprocessing
Hybrid search (combining with full-text search)
Integration with web interfaces
Security and performance considerations

The full source code, including a simple API built with FastAPI, is available on GitHub. PRs and feedback are welcome!

Resources:

Questions or feedback? Leave a comment below! 💬

DEV Community