Best ways to search in Rails, from simple to advanced:

  1. SQL ILIKE
  2. gem ransack name_or_description_cont
  3. gem pg_search - advanced Postgres search
  4. indexed search with typesense/elasticsearch/algolia
  5. AI search with embeddings

AI search lets you search by “meaning”, not by keywords!

How AI search works:

  1. create embeddings: turn texts into vectors
  2. search embeddings with gem neighbour

Install dependencies # 

# Gemfile
# pg extension to enable vectors
# https://github.com/pgvector/pgvector-ruby
gem "pgvector"
# vector search
# https://github.com/ankane/neighbor
gem "neighbor"
# create vectors
# https://github.com/crmne/ruby_llm
gem "ruby_llm"

brew install pgvector

If it doesn’t work, run

mkdir -p ~/tmp
cd ~/tmp
git clone --branch v0.8.0 https://github.com/pgvector/pgvector.git
cd pgvector

# Build and install for DocumentgreSQL 16
make USE_PGXS=1 PG_CONFIG=/opt/homebrew/opt/postgresql@16/bin/pg_config
sudo make USE_PGXS=1 PG_CONFIG=/opt/homebrew/opt/postgresql@16/bin/pg_config install

# Restart DocumentgreSQL
brew services restart postgresql@16

# config/initializers/neighbor.rb
Neighbor::PostgreSQL.initialize!

Run migrations:

rails g migration InstallNeighborVector
rails g migration AddEmbeddingToDocuments
rails g model Document title content

class InstallNeighborVector < ActiveRecord::Migration[7.1]
  def change
    enable_extension "vector"
  end
end

class AddEmbeddingToDocuments < ActiveRecord::Migration[7.1]
  def change
    add_column :documents, :embedding, :vector, limit: 1536, if_not_exists: true
  end
end

Create embeddings: # 

# app/models/document.rb
  validates :title, presence: true
  validates :content, presence: true

  has_neighbors :embedding, dimensions: 1536
  before_save :generate_embedding, if: -> { saved_change_to_title? || saved_change_to_content? }

  scope :search_by_similarity, -> (query_text) {
    query_embedding = RubyLLM.embed(query_text).vectors

    # distance: :inner_product
    nearest_neighbors(:embedding, query_embedding, distance: :cosine).first(5)
  }

  # def text_for_embedding
  #   <<~EOS
  #     Title: #{title}
  #     Content: #{content}
  #   EOS
  # end

  def generate_embedding
    text_for_embedding = [
      "Title: #{title}",
      "Content: #{content}",
    ].compact.join("\n---\n")

    begin
      self.embedding = RubyLLM.embed(text_for_embedding).vectors
    rescue RubyLLM::Error => e
    end
  end

Console

Document.create(content: "Company HR policy: Employees must...")
Document.create(content: "Company internal documentation: ...")

documents = Document.search_by_similarity("What is the company's remote work policy?")
documents.each { |document| puts "- #{document.content}" }

Controller

# app/controllers/documents_controller.rb
def index
  @documents = if params[:q].present?
    Document.all.search_by_similarity(params[:q])
  else
    Document.all
  end
end

Make it work on CI # 

# ci
-image: postgres:11-alpine
+image: pgvector/pgvector:pg16

Next steps #

  • Rate limiting; do not do typeahead search (too many requests & token$)
  • Add caching for popular queries (Query string & embedding pairs)
  • Let user see his recent searches / recently visited (might have to do fewer search queries)

Inspired by #

https://d-caponi1.medium.com/getting-set-up-with-vector-databases-in-rails-8-ac1fa2fb5b48 https://medium.com/@mauricio/how-to-add-recommendations-to-a-rails-app-with-pgvector-and-openai-881d87915fb2 https://liambx.com/blog/semantic-search-rails-neighbor-gem