sublayer 0.0.4 → 0.0.6

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: c0e73536f0855136d66096a10b35fe12acc7860e92b7c61a6408450716a7610d
-  data.tar.gz: 7de63a9367edf5e27424e15200132024e6fbde702d76f36156a0a95cb1928194
+  metadata.gz: 7a49a0cdd72c4a8171f246646ac8021f72bcdf35a0591931614cbbf7599c3bad
+  data.tar.gz: 3e0da0784fc6ace719556110c47a4145acb01d11ff90555bc25fd7ecb7c6f43f
 SHA512:
-  metadata.gz: bf794d00c4704c474586a5c3943a704109516af3dec85d431c8b610fc38d90f574a95f7c72e8d3e15daaa6214ec236d362e2350c435073807ec92c25d6cf1215
-  data.tar.gz: e5ec8d95d8966006830213a8d4058fbbf8679ae33353b41c25e5e3b620167226e760380961ae17f4adf4779d0920b7b5f60c7880c87800c1c908a3b99f763779
+  metadata.gz: 9fc567c2e6f84886b525ccac9a842b16e6e519b17d46d6eb86cec502eeafb29c110936ad8a11978124b478f8a3271415c8f96473aaebbd3d28f4678a8736340f
+  data.tar.gz: 17a53b8f40146f09f2380f4c25ad86526d6b78dac8562ab1a06f047ab05e16646cdac4bc1453533f313e6c9bdb5a6e1537a7f8efb2ab515286608eec990ee26e

data/README.md

@@ -1,16 +1,22 @@
 # Sublayer
 
-A model-agnostic Ruby Generative AI DSL and framework. Provides base classes for
+A model-agnostic Ruby AI Agent framework. Provides base classes for
 building Generators, Actions, Tasks, and Agents that can be used to build AI
 powered applications in Ruby.
 
+For more detailed documentation visit our documentation site: [https://docs.sublayer.com](https://docs.sublayer.com).
+
 ## Installation
 
 Install the gem by running the following commands:
 
-    $ bundle
-    $ gem build sublayer.gemspec
-    $ gem install sublayer-0.0.1.gem
+    $ gem install sublayer
+
+Or add this line to your application's Gemfile:
+
+```ruby
+gem 'sublayer'
+```
 
 ## Choose your AI Model
 
@@ -65,26 +71,60 @@ Sublayer.configuration.ai_provider = Sublayer::Providers::Groq
 Sublayer.configuration.ai_model = "mixtral-8x7b-32768"
 ```
 
-### Local (Experimental with Hermes2)
+### Local
 
-*Tested against the latest [Hermes 2 Mistral
-Model](https://huggingface.co/NousResearch/Hermes-2-Pro-Mistral-7B-GGUF)*
+If you've never run a local model before see the [Local Model Quickstart](#local-model-quickstart) below. Know that local models take several GB of space.
 
-Support for running a local model and serving an API on https://localhost:8080
+The model you use must have the ChatML formatted v1/chat/completions endpoint to work with sublayer (many models do by default)
 
-The simplest way to do this is to download 
-[llamafile](https://github.com/Mozilla-Ocho/llamafile) and download one of the
-server llamafiles they provide.
-
-We've also tested with [Hermes 2 Mistral](https://huggingface.co/NousResearch/Hermes-2-Pro-Mistral-7B-GGUF) from
-[Nous
-Research](https://nousresearch.com/).
+Usage:
 
+Run your local model on http://localhost:8080 and then set:
 ```ruby
 Sublayer.configuration.ai_provider = Sublayer::Providers::Local
 Sublayer.configuration.ai_model = "LLaMA_CPP"
 ```
 
+#### Local Model Quickstart:
+
+Instructions to run a local model
+
+1. Setting up Llamafile
+
+  ```bash
+  cd where/you/keep/your/projects
+  git clone git@github.com:Mozilla-Ocho/llamafile.git
+  cd llamafile
+  ```
+
+  Download: https://cosmo.zip/pub/cosmos/bin/make (windows users need this too: https://justine.lol/cosmo3/)
+
+  ```bash
+  # within llamafile directory
+  chmod +x path/to/the/downloaded/make
+  path/to/the/downloaded/make -j8
+  sudo path/to/the/downloaded/make install PREFIX=/usr/local
+  ```
+  You can now run llamfile
+
+2. Downloading Model
+
+  click [here](https://huggingface.co/NousResearch/Hermes-2-Pro-Mistral-7B-GGUF/resolve/main/Hermes-2-Pro-Mistral-7B.Q5_K_M.gguf?download=true) to download Mistral_7b.Q5_K_M (5.13 GB)
+
+3. Running Llamafile with a model
+
+  ```bash
+  llamafile -ngl 9999 -m path/to/the/downloaded/Hermes-2-Pro-Mistral-7B.Q5_K_M.gguf --host 0.0.0.0 -c 4096
+  ```
+
+  You are now running a local model on http://localhost:8080
+
+#### Recommended Settings for Apple M1 users:
+```bash
+llamafile -ngl 9999 -m Hermes-2-Pro-Mistral-7B.Q5_K_M.gguf --host 0.0.0.0 --nobrowser -c 2048 --gpu APPLE -t 12
+```
+run `sysctl -n hw.logicalcpu` to see what number to give the `-t` threads option
+
 ## Concepts
 
 ### Generators

data/lib/sublayer/agents/base.rb

@@ -0,0 +1,6 @@
+module Sublayer
+  module Agents
+    class Base
+    end
+  end
+end

data/lib/sublayer/providers/claude.rb

@@ -44,10 +44,9 @@ module Sublayer
         raise "Error generating with Claude, error: #{response.body}" unless response.code == 200
 
         text_containing_xml = JSON.parse(response.body).dig("content", 0, "text")
-        xml = text_containing_xml.match(/\<response\>(.*?)\<\/response\>/m).to_s
-        response_xml = ::Nokogiri::XML(xml)
-        function_output = response_xml.at_xpath("//response/function_calls/invoke/parameters/#{output_adapter.name}").children.to_s
+        function_output = Nokogiri::HTML.parse(text_containing_xml.match(/\<#{output_adapter.name}\>(.*?)\<\/#{output_adapter.name}\>/m)[1]).text
 
+        raise "Claude did not format response, error: #{response.body}" unless function_output
         return function_output
       end
     end

data/lib/sublayer/providers/gemini.rb

@@ -5,21 +5,49 @@ module Sublayer
   module Providers
     class Gemini
       def self.call(prompt:, output_adapter:)
+        system_prompt = <<-PROMPT
+        You have access to a set of tools to answer the prompt.
+
+        You may call tools like this:
+        <tool_calls>
+          <tool_call>
+            <tool_name>$TOOL_NAME</tool_name>
+            <parameters>
+              <$PARAMETER_NAME>$VALUE</$PARAMETER_NAME>
+              ...
+            </parameters>
+          </tool_call>
+        </tool_calls>
+
+        Here are the tools available:
+        <tools>
+          <tool>
+            #{output_adapter.to_xml}
+          </tool>
+        </tools>
+
+        Respond only with valid xml.
+        The entire response should be wrapped in a <response> tag.
+        Your response should call a tool inside a <tool_calls> tag.
+        PROMPT
+
         response = HTTParty.post(
           "https://generativelanguage.googleapis.com/v1beta/models/#{Sublayer.configuration.ai_model}:generateContent?key=#{ENV['GEMINI_API_KEY']}",
           body: {
-            tools: { function_declarations: [output_adapter.to_hash] },
-            contents: { role: "user", parts: { text: prompt } }
+            contents: { role: "user", parts: { text: "#{system_prompt}\n#{prompt}" } }
           }.to_json,
           headers: {
             "Content-Type" => "application/json"
-          })
+          }
+        )
+
+        raise "Error generating with Gemini, error: #{response.body}" unless response.success?
 
-        part = response.dig('candidates', 0, 'content', 'parts', 0)
-        raise "No function called" unless part['functionCall']
+        text_containing_xml = response.dig('candidates', 0, 'content', 'parts', 0, 'text')
+        tool_output = Nokogiri::HTML.parse(text_containing_xml.match(/\<#{output_adapter.name}\>(.*?)\<\/#{output_adapter.name}\>/m)[1]).text
 
-        args = part['functionCall']['args']
-        args[output_adapter.name]
+        raise "Gemini did not format response, error: #{response.body}" unless tool_output
+        return tool_output
       end
     end
   end

data/lib/sublayer/providers/groq.rb

@@ -6,27 +6,27 @@ module Sublayer
     class Groq
       def self.call(prompt:, output_adapter:)
         system_prompt = <<-PROMPT
-        In this environment you have access to a set of tools you can use to answer the user's question.
+        You have access to a set of tools to answer the prompt.
 
-        You may call them like this:
-        <function_calls>
-        <invoke>
-          <tool_name>$TOOL_NAME</tool_name>
-          <parameters>
-          <#{output_adapter.name}>value</#{output_adapter.name}>
-          ...
-          </parameters>
-        </invoke>
-        </function_calls>
+        You may call tools like this:
+        <tool_calls>
+          <tool_call>
+            <tool_name>$TOOL_NAME</tool_name>
+            <parameters>
+              <#{output_adapter.name}>$VALUE</#{output_adapter.name}>
+              ...
+            </parameters>
+          </tool_call>
+        </tool_calls>
 
         Here are the tools available:
         <tools>
-        #{output_adapter.to_xml}
+          #{output_adapter.to_xml}
         </tools>
 
         Respond only with valid xml.
         The entire response should be wrapped in a <response> tag.
-        Any additional information not inside a tool call should go in a <scratch> tag.
+        Your response should call a tool inside a <tool_calls> tag.
         PROMPT
 
         response = HTTParty.post(
@@ -41,12 +41,11 @@ module Sublayer
           }.to_json
         )
 
-        text_containing_xml = JSON.parse(response.body).dig("choices", 0, "message", "content")
-        xml = text_containing_xml.match(/\<response\>(.*?)\<\/response\>/m).to_s
-        response_xml = ::Nokogiri::XML(xml)
-        function_output = response_xml.at_xpath("//response/function_calls/invoke/parameters/command").children.to_s
+        text_containing_xml = response.dig("choices", 0, "message", "content")
+        tool_output = Nokogiri::HTML.parse(text_containing_xml.match(/\<#{output_adapter.name}\>(.*?)\<\/#{output_adapter.name}\>/m)[1]).text
+        raise "Groq did not format response correctly, error: #{response.body}" unless tool_output
 
-        return function_output
+        return tool_output
       end
     end
   end

data/lib/sublayer/providers/local.rb

@@ -6,23 +6,23 @@ module Sublayer
     class Local
       def self.call(prompt:, output_adapter:)
         system_prompt = <<-PROMPT
-        You are a function calling AI agent
-        You can call only one function at a time
-        You are provided with function signatures within <tools></tools> XML tags.
+        You have access to a set of tools to respond to the prompt.
 
-        Please call a function and wait for results to be provided to you in the next iteration.
-        Don't make assumptions about what values to plug into function arguments.
+        You may call a tool with xml like this:
+        <parameters>
+          <#{output_adapter.name}>$VALUE</#{output_adapter.name}>
+          ...
+        </parameters>
 
-        Here are the available tools:
+        Here are descriptions of the available tools:
         <tools>
-        #{output_adapter.to_hash.to_json}
+          <tool>
+            #{output_adapter.to_xml}
+          </tool>
         </tools>
 
-        For the function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags as follows:
-
-        <tool_call>
-        {"arguments": <args-dict>, "name": <function-name>}
-        </tool_call>
+        Respond only with valid xml.
+        Your response should call a tool with xml inside a <parameters> tag.
         PROMPT
 
         response = HTTParty.post(
@@ -34,17 +34,16 @@ module Sublayer
           body: {
             "model": Sublayer.configuration.ai_model,
             "messages": [
-              { "role": "system", "content": system_prompt },
-              { "role": "user", "content": prompt }
+              { "role": "user", "content": "#{system_prompt}\n#{prompt}}" }
             ]
           }.to_json
         )
 
-        text_containing_xml = JSON.parse(response.body).dig("choices", 0, "message", "content")
-        results = JSON.parse(::Nokogiri::XML(text_containing_xml).at_xpath("//tool_call").children.to_s.strip)
-        function_output = results["arguments"][output_adapter.name]
+        text_containing_xml = response.dig("choices", 0, "message", "content")
+        tool_output = Nokogiri::HTML.parse(text_containing_xml.match(/\<#{output_adapter.name}\>(.*?)\<\/#{output_adapter.name}\>/m)[1]).text
+        raise "The response was not formatted correctly: #{response.body}" unless tool_output
 
-        return function_output
+        return tool_output
       end
     end
   end

data/lib/sublayer/tasks/base.rb

@@ -0,0 +1,6 @@
+module Sublayer
+  module Tasks
+    class Base
+    end
+  end
+end

data/lib/sublayer/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module Sublayer
-  VERSION = "0.0.4"
+  VERSION = "0.0.6"
 end

data/sublayer.gemspec

@@ -11,10 +11,12 @@ Gem::Specification.new do |spec|
 
   spec.summary = "A model-agnostic Ruby GenerativeAI DSL and Framework"
   spec.description = "A DSL and framework for building AI powered applications through the use of Generators, Actions, Tasks, and Agents"
-  spec.homepage = "https://www.sublayer.com"
+  spec.homepage = "https://docs.sublayer.com"
   spec.required_ruby_version = ">= 2.6.0"
 
-  spec.metadata["homepage_uri"] = spec.homepage
+  spec.metadata["homepage_uri"] = "https://docs.sublayer.com"
+  spec.metadata["documentation_uri"] = "https://docs.sublayer.com"
+  spec.metadata["bug_tracker_uri"] = "https://github.com/sublayerapp/sublayer/issues"
 
   # Specify which files should be added to the gem when it is released.
   # The `git ls-files -z` loads the files in the RubyGem that have been added into git.

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: sublayer
 version: !ruby/object:Gem::Version
-  version: 0.0.4
+  version: 0.0.6
 platform: ruby
 authors:
 - Scott Werner
 autorequire:
 bindir: bin
 cert_chain: []
-date: 2024-04-04 00:00:00.000000000 Z
+date: 2024-04-14 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: ruby-openai
@@ -153,6 +153,7 @@ files:
 - examples/invalid_to_valid_json_generator.rb
 - lib/sublayer.rb
 - lib/sublayer/actions/base.rb
+- lib/sublayer/agents/base.rb
 - lib/sublayer/components/output_adapters.rb
 - lib/sublayer/components/output_adapters/single_string.rb
 - lib/sublayer/generators/base.rb
@@ -161,13 +162,16 @@ files:
 - lib/sublayer/providers/groq.rb
 - lib/sublayer/providers/local.rb
 - lib/sublayer/providers/open_ai.rb
+- lib/sublayer/tasks/base.rb
 - lib/sublayer/version.rb
 - sublayer.gemspec
-homepage: https://www.sublayer.com
+homepage: https://docs.sublayer.com
 licenses:
 - MIT
 metadata:
-  homepage_uri: https://www.sublayer.com
+  homepage_uri: https://docs.sublayer.com
+  documentation_uri: https://docs.sublayer.com
+  bug_tracker_uri: https://github.com/sublayerapp/sublayer/issues
 post_install_message:
 rdoc_options: []
 require_paths: