💎 Rag Embeddings

rag_embeddings is a native Ruby library for efficient storage and comparison of AI-generated embedding vectors (float arrays) using high-performance C extensions. It is designed for seamless integration with external LLMs (Ollama, OpenAI, Mistral, etc) and works perfectly for RAG (Retrieval-Augmented Generation) applications.

• C extension for maximum speed in cosine similarity and vector allocation
• Compatible with langchainrb for embedding generation (Ollama, OpenAI, etc)
• SQLite-based storage with vector search capabilities
• RSpec tested

• Creation of embedding objects from LLM-generated float arrays
• Cosine similarity calculation in C for speed and safety
• Embedding + text storage in SQLite (BLOB)
• Retrieve top-K most similar texts to a query using cosine similarity
• Memory-safe and 100% Ruby compatible
• Plug-and-play for RAG, semantic search, and retrieval AI

• Question Answering over Documents: Instantly search and retrieve the most relevant document snippets from thousands of articles, FAQs, or customer support logs in your Ruby app.
• Semantic Search for E-commerce: Power product search with semantic understanding, returning items similar in meaning, not just keywords.
• Personalized Recommendations: Find related content (articles, products, videos) by comparing user preferences and content embeddings.
• Knowledge Base Augmentation: Use with OpenAI or Ollama to enhance chatbots, letting them ground answers in your company’s internal documentation or wiki.
• Fast Prototyping for AI Products: Effortlessly build MVPs for RAG-enabled chatbots, semantic search tools, and AI-driven discovery apps—all in native Ruby.

• Ruby >= 3.3
• langchainrb (for embedding)
• At the moment ollama is used as LLM so it must be active and working, although there are some workarounds
• sqlite3 (for storage)

Requires a working C compiler in order to build the native extension

gem install rag_embeddings

If you'd rather install it using bundler, add a line for it in your Gemfile (but set the require option to false, as it is a standalone tool):

gem "rag_embeddings", require: false

require "rag_embeddings"
embedding = RagEmbeddings.embed("Hello world, this is RAG!")
# embedding is a float array

The default model is llama3.2 but you can set another one (reload the console as the llm is memoized):

embedding = RagEmbeddings.embed("Hello world, this is RAG!", model: 'qwen3:0.6b')

c_embedding = RagEmbeddings::Embedding.from_array(embedding)
puts "Dimension: #{c_embedding.dim}"
# Dimension: 1024 # qwen3:0.6b
# Dimension: 3072 # llama3.2

puts "Ruby array: #{c_embedding.to_a.inspect}"

emb1 = RagEmbeddings.embed("Hello world!")
emb2 = RagEmbeddings.embed("Hi universe!")
obj1 = RagEmbeddings::Embedding.from_array(emb1)
obj2 = RagEmbeddings::Embedding.from_array(emb2)
sim = obj1.cosine_similarity(obj2)
puts "Cosine similarity: #{sim}"  # Value between -1 and 1

db = RagEmbeddings::Database.new("embeddings.db")
db.insert("Hello world!", RagEmbeddings.embed("Hello world!"))
db.insert("Completely different sentence", RagEmbeddings.embed("Completely different sentence"))

# Find the most similar text to a query
result = db.top_k_similar("Hello!", k: 1)
puts "Most similar text: #{result.first[1]}, score: #{result.first[2]}"

# load all .txt files
files = Dir["./docs/*.txt"].map { |f| [File.basename(f), File.read(f)] }

db = RagEmbeddings::Database.new("knowledge_base.db")
files.each do |name, text|
  vector = RagEmbeddings.embed(text)
  db.insert(name, vector)
end

puts "Indexed #{files.size} documents."

require "openai"        # or your favorite LLM client

# 1) build or open your vector store
db = RagEmbeddings::Database.new("knowledge_base.db")

# 2) embed your user question
client      = OpenAI::Client.new(api_key: ENV.fetch("OPENAI_API_KEY"))
q_embedding = client.embeddings(
  parameters: {
    model: "text-embedding-ada-002",
    input: "What are the benefits of retrieval-augmented generation?"
  }
).dig("data", 0, "embedding")

# 3) retrieve top-3 relevant passages
results = db.top_k_similar(q_embedding, k: 3)

# 4) build a prompt for your LLM
context = results.map { |id, text, score| text }.join("\n\n---\n\n")
prompt  = <<~PROMPT
  You are an expert.  
  Use the following context to answer the question:

  CONTEXT:
  #{context}

  QUESTION:
  What are the benefits of retrieval-augmented generation?
PROMPT

# 5) call the LLM for final answer
response = client.chat(
  parameters: {
    model: "gpt-4o",
    messages: [{ role: "user", content: prompt }]
  }
)
puts response.dig("choices", 0, "message", "content")

# use SQLite :memory: for ephemeral experiments
db = RagEmbeddings::Database.new(":memory:")

# insert & search exactly as with a file-backed DB
db.insert("Quick test", RagEmbeddings.embed("Quick test"))
db.top_k_similar("Test", k: 1)

rag_embeddings combines the simplicity of Ruby with the performance of C to deliver fast vector operations for RAG applications.

The library uses a hybrid memory-storage approach:

1. In-Memory Processing: All vector operations (cosine similarity calculations, embedding manipulations) happen entirely in memory using optimized C code
2. Persistent Storage: SQLite serves as a simple, portable storage layer for embeddings and associated text
3. Dynamic C Objects: Embeddings are managed as native C structures with automatic memory management

C Extension (embedding.c)

• Handles all computationally intensive operations
• Manages dynamic vector dimensions (adapts to any LLM output size)
• Performs cosine similarity calculations with optimized algorithms
• Ensures memory-safe operations with proper garbage collection integration

Ruby Interface

• Provides an intuitive API for vector operations
• Integrates seamlessly with LLM providers via langchainrb
• Handles database operations and query orchestration

SQLite Storage

• Stores embeddings as BLOBs alongside their associated text
• Provides persistent storage without requiring external databases
• Supports both file-based and in-memory (:memory:) databases
• Enables portable, self-contained applications

1. Text → Embedding: Generate vectors using your preferred LLM (Ollama, OpenAI, etc.)
2. Memory Allocation: Create C embedding objects with Embedding.from_array()
3. Storage: Persist embeddings and text to SQLite for later retrieval
4. Query Processing:
   • Load query embedding into memory
   • Compare against stored embeddings using fast C-based cosine similarity
   • Return top-K most similar results ranked by similarity score

Performance: Critical operations run in optimized C code, delivering significant speed improvements over pure Ruby implementations.

Memory Efficiency: While embeddings are stored in SQLite, all vector computations happen in memory, avoiding I/O bottlenecks during similarity calculations.

Simplicity: SQLite eliminates the need for complex vector database setups while maintaining good performance for moderate-scale applications.

Portability: The entire knowledge base fits in a single SQLite file, making deployment and backup trivial.

• Embedding creation: ~82ms for 10,000 operations
• Cosine similarity: ~107ms for 10,000 calculations
• Memory usage: ~34MB for 10,000 embeddings
• Scalability: Suitable for thousands to tens of thousands of vectors

For applications requiring millions of vectors, consider specialized vector databases (Faiss, sqlite-vss) while using this library for prototyping and smaller-scale production use.

• Embedding provider: switch model/provider in engine.rb (Ollama, OpenAI, etc)
• Database: set the SQLite file path as desired

If you need to customize the c part (ext/rag_embeddings/embedding.c), recompile it with:

rake compile

To run all specs (RSpec required):

bundle exec rspec

bundle exec rspec spec/performance_spec.rb

You'll get something like this in random order:

Performance test with embedding size: 768
Embedding creation (10000 times): 19 ms
Cosine similarity (10000 times): 27 ms
RSS: 132.3 MB

Memory usage test with embedding size: 768
Memory usage delta: 3.72 MB for 10000 embeddings


Performance test with embedding size: 2048
Embedding creation (10000 times): 69 ms
Cosine similarity (10000 times): 73 ms
RSS: 170.08 MB

Memory usage test with embedding size: 2048
Memory usage delta: 25.11 MB for 10000 embeddings


Performance test with embedding size: 3072
Embedding creation (10000 times): 98 ms
Cosine similarity (10000 times): 112 ms
RSS: 232.97 MB

Memory usage test with embedding size: 3072
Memory usage delta: 60.5 MB for 10000 embeddings


Performance test with embedding size: 4096
Embedding creation (10000 times): 96 ms
Cosine similarity (10000 times): 140 ms
RSS: 275.2 MB

Memory usage test with embedding size: 4096
Memory usage delta: 92.41 MB for 10000 embeddings

Open an issue or contact the maintainer for questions, suggestions, or bugs.

Happy RAG! 🚀