Load the gem outside Rails with Zeitwerk

CHANGELOG.md

@@ -15,6 +15,7 @@
 - [FEATURE] [https://github.com/patterns-ai-core/langchainrb/pull/1085] Add `Langchain::Tool::SearchApi`, a wrapper around [SearchApi](https://www.searchapi.io) to search Google, Bing and other search engines.
 - [BUGFIX] [https://github.com/patterns-ai-core/langchainrb/pull/1083] `Langchain::LLM::Response::AnthropicResponse#prompt_tokens` (and `#total_tokens`) include the prompt-cache read and creation tokens that Anthropic reports apart from `input_tokens`; they were left out before.
 - [BUGFIX] [https://github.com/patterns-ai-core/langchainrb/pull/1084] `Langchain::LLM::Response::GoogleGeminiResponse#completion_tokens` includes the thinking tokens (`thoughtsTokenCount`), so prompt plus completion tokens match `total_tokens` for thinking models.
+- [BUGFIX] [https://github.com/patterns-ai-core/langchainrb/pull/1080] `require "langchain"` works outside Rails again: without `Rails::Engine` the gem sets up Zeitwerk itself. Since the move to a Rails engine it raised `NameError` (uninitialized constant Rails) in a plain Ruby script.
 - [BUGFIX] [https://github.com/patterns-ai-core/langchainrb/pull/1079] `Langchain::DependencyHelper#depends_on` accepts a gem that's in the bundle through another gem's gemspec instead of the `Gemfile`, such as `aws-sdk-bedrockruntime` declared by a gem that depends on langchainrb; it raised `LoadError` before.
 - [FEATURE] [https://github.com/patterns-ai-core/langchainrb/pull/1078] Add `Langchain::Vectorsearch::Milvus#update_texts`; with the default `auto_id` schema Milvus gives each updated text a new id, returned under `"data" => "upsertIds"`.
 - [FEATURE] [https://github.com/patterns-ai-core/langchainrb/pull/1074] Add `Langchain::Vectorsearch::Pinecone#remove_texts`.

lib/langchain.rb

@@ -1,6 +1,33 @@
 require "logger"
 require "langchain/version"
-require "langchain/engine"
+if defined?(Rails::Engine)
+  require "langchain/engine"
+else
+  # Outside a Rails app there's no engine to autoload the gem, so set up Zeitwerk with the
+  # same acronyms as config/initializers/inflections.rb
+  require "zeitwerk"
+
+  loader = Zeitwerk::Loader.for_gem
+  loader.ignore("#{__dir__}/langchainrb.rb", "#{__dir__}/langchain/engine.rb", "#{__dir__}/langchain/application_record.rb")
+  loader.inflector.inflect(
+    "ai21" => "AI21",
+    "ai21_response" => "AI21Response",
+    "csv" => "CSV",
+    "google_vertexai" => "GoogleVertexAI",
+    "html" => "HTML",
+    "json" => "JSON",
+    "jsonl" => "JSONL",
+    "llm" => "LLM",
+    "mistralai" => "MistralAI",
+    "mistralai_message" => "MistralAIMessage",
+    "mistralai_response" => "MistralAIResponse",
+    "openai" => "OpenAI",
+    "openai_message" => "OpenAIMessage",
+    "openai_response" => "OpenAIResponse",
+    "pdf" => "PDF"
+  )
+  loader.setup
+end
 
 module Langchain
   class << self

spec/langchain_spec.rb

@@ -1,7 +1,22 @@
 # frozen_string_literal: true
 
+require "open3"
+
 RSpec.describe Langchain do
   it "has a version number" do
     expect(Langchain::VERSION).not_to be nil
   end
+
+  it "loads every class without Rails" do
+    script = <<~RUBY
+      require "langchain"
+      abort "Rails was loaded" if defined?(Rails)
+      Zeitwerk::Loader.eager_load_all
+      print Langchain::LLM::OpenAI.name, " ", Langchain::LLM::Response::OpenAIResponse.name
+    RUBY
+    output, status = Open3.capture2e(RbConfig.ruby, "-I", Langchain.root.to_s, "-e", script)
+
+    expect(status).to be_success, output
+    expect(output).to eq("Langchain::LLM::OpenAI Langchain::LLM::Response::OpenAIResponse")
+  end
 end