Top Gems for AI Development in Ruby on Rails

In this post, I’ll walk through what I consider the top gems for AI in Rails (as of October 2025), what they’re good for, pros and trade-offs, and tips for using them effectively in real applications.

The ruby-openai gem is often the first stop for Rails developers curious about AI. It provides a Ruby wrapper around the OpenAI REST API, supporting chat completions, embeddings, image generation, streaming, Whisper (audio) and more.

It’s the de facto choice for basic AI operations in Rails, sending messages to GPT, retrieving embeddings, generating images, or transcribing audio. Because it’s broad in scope, it handles many use cases you'll come across when prototyping.

A frequent complaint is that streaming support is barebones or requires manual work. The gem doesn’t include a full streaming conversation library by default and that you may need to roll your own orchestration using SSE or chunked responses. Also, because it's a general wrapper, it doesn’t enforce patterns like chaining or retrieval-augmented generation (RAG). As your app grows, you may find yourself reworking request plumbing.

While ruby-openai is widely used, the official OpenAI Ruby SDK is gaining traction for its tighter alignment with OpenAI’s latest APIs.

Because it’s official, it often gets early support for new endpoints or features (e.g. streaming, updates). It comes with strong typing, documentation, and native support for the latest OpenAI primitives. It’s more focused on OpenAI services than being a full AI framework. If you want to switch LLM providers or build complex pipelines, you may quickly hit its limits. Also, community integrations into Rails (like for embeddings or agent frameworks) are still catching up.

If you want structured AI workflows, chaining multiple prompts, integrating search, or building agents, langchainrb is the go-to Ruby port of the popular LangChain framework.

It lets you glue together LLM calls, store memory, embed with vector databases, and build simple agent systems. For example, you can build a question-answer interface that first retrieves relevant context via embeddings and then passes that context into a generation call (RAG) rather than sending raw prompts.

You can also integrate with its Rails companion langchainrb_rails, which provides generators to scaffold vectorsearch support in your models.

It has a learning curve. The abstractions can feel heavy if all you want is a simple chat integration. Configuration quirks sometimes confuse newcomers, but for apps that grow in AI complexity, it pays dividends by reducing boilerplate and enforcing structure.

AI embeddings are powerful only when your database can handle them. Enter neighbor, a gem that makes working with vectors in Rails and Postgres smoother. It wraps around pgvector (or Postgres vector support) to integrate embeddings into ActiveRecord models.

It handles migrations, schema dumping, correct index types (like hnsw), and vector queries so you don’t have to write raw SQL or battle schema.rb errors for unknown types. In cases without neighbor, applications with vector columns sometimes fail schema dumps or require custom SQL in migrations.

Because the vector world is still maturing in Postgres, edge cases appear when your embedding dimension changes or you adjust vector algorithms. Also, neighbor doesn’t itself produce embeddings, you still need something like ruby-openai or langchainrb to generate them.

This is a lightweight gem designed specifically for chat completion workflows, focusing on structure and output formats.

If your main use case is building chat or conversational UI's, this gem is handy because it keeps the API surface narrow and structured. It can simplify handling replies, roles, and result parsing.

Because it’s narrowly scoped, you’ll likely mix it with other tools for embeddings, memory, or agent logic. It’s best suited when your AI feature is primarily chat-based and you don’t need full pipeline flexibility.

Working with AI in Ruby on Rails today feels a bit like those early days of Rails itself, fast-moving, a bit rough in places, but full of potential. The tooling has matured quickly over the past year, and it’s finally at the point where things mostly just work. You can install a gem, make a call to an LLM, and have something useful running in an afternoon.

That said, there are still moments that remind you how new this all is. API limits, occasional cryptic errors, and unpredictable model behaviour can throw you off if you’re expecting perfection. But those moments are becoming fewer and easier to work around. What’s exciting is how easily AI now fits into the Rails mindset, convention over configuration, fast feedback loops, and sensible defaults.

For me, the sweet spot is using these tools to quietly improve existing apps rather than building something that screams AI. Automating summaries, improving search, giving users helpful nudges, all the subtle things that make an app feel more polished. That’s where Rails and AI overlap beautifully: pragmatic solutions that make everyday development a little smarter.

If you’re thinking about dipping your toes in, start small. Pick one gem, ruby-openai or langchainrb, and build something fun that solves a problem you actually have. You’ll quickly see how natural it feels to work with these tools inside Rails, and that’s when the ideas start flowing.