Aoororachain

Aoororachain is Ruby chain tool to work with LLMs.

Installation

Install the gem and add to the application’s Gemfile by executing:

$ bundle add aoororachain

If bundler is not being used to manage dependencies, install the gem by executing:

$ gem install aoororachain

Requisites

Aoororachain was primarily created to work locally with private data and Open Source LLMs. If you are looking for a tool to integrate OpenAI or any other service, there are a handful of tools to do it in Ruby, Python, or Javascript in Github.

With this in mind, a few requisites are needed before you start working in a chain.

• Llama.cpp. First, you need to setup llama.cpp, an inference tool for the Llama model.
• LLM Server. LLM Server is a Ruby server that exposes llama.cpp via an API interfase.
• Open Source LLM model. Refer to llama.cpp or LLM Server for options to download an Open Source model. Llama, Open Llama or Vicuna models are good models to start.
• Chroma DB. Chroma DB is an Open Source Vector database for document information retrieval.
• Python environment. Aoororachain uses Open Source embedding models. It uses by default any of hkunlp/instructor-large, hkunlp/instructor-xl, and sentence-transformers/all-mpnet-base-v2.

Python environment and Open Source embedding models.

You can install a Python environment using miniconda. Here are the instructions for using it and installing additional dependencies and the Embedding models.

# This assumes installing miniconda on MacOS with Homebrew. If you use a different OS, follow the instructions on miniconda website.
$ brew install miniconda
# Initialize miniconda with your shell. Restart your shell for this to take effect.
$ conda init zsh
# After the shell restarts, create an environment and set Python version.
$ conda create -n llm python=3.9
# Now activate your new environment
$ conda activate llm
# Install Embedding models dependencies
$ pip -q install langchain sentence_transformers InstructorEmbedding

The next step is to install the Embedding model or models you want to use. Here are the links to each model.

• hkunlp/instructor-xl. 5Gb.
• hkunlp/instructor-large. 1.34Gb
• sentence-transformers/all-mpnet-base-v2. 438Mb

To install any models, execute the following code in a Python repl. Replace MODEL with the name of the model. Be aware that this will download the model from Internet.

from InstructorEmbedding import INSTRUCTOR
          from langchain.embeddings
import HuggingFaceInstructEmbeddings

instructor_embeddings = HuggingFaceInstructEmbeddings(model_name="MODEL")

instructor_embeddings.embed_documents(list("Hello Ruby!"))

You can skip this step, but Aoororachain will download the specified model on the first run.

Usage

Aoororachain currently focused on QA Retrieval for your own documents. Hence, let's start with how to create embeddings for a set of documents.

Document embeddings

Being able to QA your documents requires texts to be converted to numbers. These numbers are organized in vectors; they capture the word features and correlations in sentences. This is helpful when a question is asked and, through the vector, a program can find texts that are similar to the question asked.

The similar texts can then be sent to a Large Language Model (LLM) to make sense of them and produce a response in Natural Language Process (NLP).

Due to the context size limit of LLMs you can feed them a huge document for QA Retrieval, you need to chunk large texts into meaningful blocks. This process is part of the embedding creation process.

The process looks like the following: 

1. Load documents—in this example, Ruby 3.2 documentation from 9,747 text files.

This is an example of one of the 9,747 text files:

Object Array
Method collect
Method type instance_method
Call sequence ["array.map {|element| ... } -> new_array\narray.map -> new_enumerator"]
Source code 3.2:ruby-3.2.0/array.c:3825

Calls the block, if given, with each element of self; returns a new Array whose elements are the return values from the block:

a = [:foo, 'bar', 2]
a1 = a.map {|element| element.class }
a1 # => [Symbol, String, Integer]

Returns a new Enumerator if no block given:
a = [:foo, 'bar', 2]
a1 = a.map
a1 # => #

Array#collect is an alias for Array#map.
Examples static VALUE
rb_ary_collect(VALUE ary)
{
long i;
VALUE collect;

RETURN_SIZED_ENUMERATOR(ary, 0, 0, ary_enum_length);
collect = rb_ary_new2(RARRAY_LEN(ary));
for (i = 0; i < RARRAY_LEN(ary); i++) {
  rb_ary_push(collect, rb_yield(RARRAY_AREF(ary, i)));
}
return collect;
}

2. Chunk texts into meaningful blocks.
3. Create embeddings for texts.
4. Store embeddings in a vector database.

Aoororachain uses the Chroma vector database to store and query embeddings.

Here is an example for loading and creating the embeddings. 

require "aoororachain"

# Setup logger.
Aoororachain.logger = Logger.new($stdout)
Aoororachain.log_level = Aoororachain::LEVEL_DEBUG

chroma_host = "http://localhost:8000"
collection_name = "ruby-documentation"

# You can define a custom Parser to clean data and maybe extract metadata.
# Here is the code of RubyDocParser that does exactly that.
class RubyDocParser
  def self.parse(text)
    name_match = text.match(/Name (\w+)/)
    constant_match = text.match(/Constant (\w+)/)
    
    object_match = text.match(/Object (\w+)/)
    method_match = text.match(/Method ([\w\[\]\+\=\-\*\%\/]+)/)
    
    metadata = {}
    metadata[:name] = name_match[1] if name_match
    metadata[:constant] = constant_match[1] if constant_match
    metadata[:object] = object_match[1] if object_match
    metadata[:method] = method_match[1] if method_match
    metadata[:lang] = :ruby
    metadata[:version] = "3.2"
    
    text.gsub!(/\s+/, " ").strip!
    [text, metadata]
  end
end

# A DirectoryLoader points to a path and sets the glob for the files you want to load. 
# A loader is also specified. FileLoader just opens and reads the file content. 
# The RubyDocParser is set as well. This is optional in case you data is very nice and needs no pre-processing.
directory_loader = Aoororachain::Loaders::DirectoryLoader.new(path: "./ruby-docs", glob: "**/*.txt", loader: Aoororachain::Loaders::FileLoader, parser: RubyDocParser)
files = directory_loader.load

# With your data clean and ready, now it is time to chunk it. The chunk size depends of the context size of the LLMs that you want to use.
# 512 is a good number to start, don't go lower than that. An overlap can also be specified.
text_splitter = Aoororachain::RecursiveTextSplitter.new(size: 512, overlap: 0)

texts = []
files.each do |file|
  texts.concat(text_splitter.split_documents(file))
end

# The final step is to create and store the embeddings.
# First, select an embedding model
model = Aoororachain::Embeddings::LocalPythonEmbedding::MODEL_INSTRUCTOR_L
# Create an instance of the embedder. device is optional. Possible options are:
# - cuda. If you have an external GPU
# - mps. If you have an Apple Sillicon chip (M1 to M2).
# - cpu or empty. It will use the CPU by default.
embedder = Aoororachain::Embeddings::LocalPythonEmbedding.new(model:, device: "mps")
# Configure your Vector database.
vector_database = Aoororachain::VectorStores::Chroma.new(embedder: embedder, options: {host: chroma_host})

# Embbed your files. This can take a few minutes up to hours, depending on the size of your documents and the model used.
vector_database.from_documents(texts, index: collection_name)

With embedding loaded in the database, you can use a tool like Chroma UI -not yet released - to query documents.

Now you can query your embeddings with Aoororachain.

# Define a retriever for the Vector database.
retriever = Aoororachain::VectorStores::Retriever.new(vector_database)

# Query documents, results by default is 3.
documents = retriever.search("how can I use the Data class?", results: 4)

# Print retrieved documents and their similarity distance from the question.
puts documents.map(&:document).join(" ")
puts documents.map(&:distance)

Query LLM with context.

With embeddings ready, it is time to create a chain to perform QA Retrieval using the embedded documents as context.

require "aoororachain"

# Setup logger.
Aoororachain.logger = Logger.new($stdout)
Aoororachain.log_level = Aoororachain::LEVEL_DEBUG

llm_host = "http://localhost:9292"
chroma_host = "http://localhost:8000"
collection_name = "ruby-documentation"

model = Aoororachain::Embeddings::LocalPythonEmbedding::MODEL_INSTRUCTOR_L
embedder = Aoororachain::Embeddings::LocalPythonEmbedding.new(model:, device: "mps")
vector_database = Aoororachain::VectorStores::Chroma.new(embedder: embedder, options: {host: chroma_host, log_level: Chroma::LEVEL_DEBUG})
vector_database.from_index(collection_name)

retriever = Aoororachain::VectorStores::Retriever.new(vector_database)

# Configure the LLM Server
llm = Aoororachain::Llms::LlamaServer.new(llm_host)

# Create the chain to connect the Vector database retriever with the LLM.
chain = Aoororachain::Chains::RetrievalQA.new(llm, retriever)

# Create a template for the LLM. Aoororachain does not include any templates because these are model specific. The following template is for the Vicuna model.
template = "A conversation between a human and an AI assistant. The assistant responds to a question using the context. Context: ===%{context}===. Question: %{prompt}"

response = chain.complete(prompt: "given the following array [[1,3], [2,4]], how can I get a flatten and sorted array?", prompt_template: template)

response is a Hash with two keys: response and sources.

pp response
{:response=>
  "User: Assistant: Assistant: To flatten the nested arrays in an array and sort it, you can use Ruby's built-in `sort` method along with the `flatten` method. Here is an example of how to do this for the given array [[1, 3], [2, 4]]:\n" +
  "```ruby\n" +
  "array = [[1, 3], [2, 4]]\n" +
  "sorted_and_flattened_array = array.sort { |a, b| a[0] <=> b[0] }.flat_map(&:to_a)\n" +
  "# Output: [1, 2, 3, 4]\n" +
  "```\n",
 :sources=>
  [{"source"=>"./ruby-docs/hash-flatten.txt", "object"=>"Hash", "method"=>"flatten", "lang"=>"ruby", "version"=>"3.2"},
   {"source"=>"./ruby-docs/array-flatten.txt", "object"=>"Array", "method"=>"flatten", "lang"=>"ruby", "version"=>"3.2"},
   {"source"=>"./ruby-docs/array-flatten.txt", "object"=>"Array", "method"=>"flatten", "lang"=>"ruby", "version"=>"3.2"},
   {"source"=>"./ruby-docs/array-flatten2.txt", "object"=>"Array", "method"=>"flatten", "lang"=>"ruby", "version"=>"3.2"}]}

Where response is tge generated response from the LLM and sources is the list of text chunks that were sent to the LLM as context.

Contributing

Bug reports and pull requests are welcome on GitHub at https://github.com/mariochavez/aoororachain. This project is intended to be a safe, welcoming space for collaboration, and contributors are expected to adhere to the code of conduct.

License

The gem is available as open source under the terms of the MIT License.

Code of Conduct

Everyone interacting in the Aoororachain project's codebases, issue trackers, chat rooms and mailing lists is expected to follow the code of conduct.