llm_chain 0.3.0 → 0.4.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: efda7d777bfd7333f75b65ccc74d46abaff6c70e2a30a31709988811ed54cfff
-  data.tar.gz: 4ea72695094eff0ee6d1a1acc029f66ea34be3a0052d39d33d70a9c6f5a1e1d7
+  metadata.gz: 8edbb57db3e5fe7b44c7e4ebec1e026dea81bc1e30fc140e9ac35549b5c862e8
+  data.tar.gz: 2c1dd32185e96ea57bdae128934726b556bfbaaa8a913cb0062d53021cf7df2f
 SHA512:
-  metadata.gz: 94423d0dc1cc4d6305aeb94dfc06fe50882726df1d1f72d03596317db4c37cb517fc9ad272236717439878fac7ca46d23354911e64bd1801b9e250d61d237a4b
-  data.tar.gz: a1ca64366d7343ae2ad8622372d19c97291a0046cf7056af0e9f79902c31e113d8dac8008e687ee53c9bd32f5f24243ded5b74d7dd256f6c23ba9aedb319862d
+  metadata.gz: aa19f25581b568ca3cd2f5d9569afd0f8beab706f56ce63b043a6bd81965a9386efb4b27aff0563bc9530576aaa836a5db092864f852962c47ea5f19fb859880
+  data.tar.gz: 4d38b4daf714d28b298a41c555add9739faee9cdf7979de52b3ff944f2ed275694aa597f1faafed21fb844002b1915591ef043bcd6686483d931a59128ca62e7

data/README.md

@@ -46,7 +46,9 @@ basic example:
 require 'llm_chain'
 
 memory = LLMChain::Memory::Array.new(max_size: 1)
-chain = LLMChain::Chain.new(model: "qwen3:1.7b", memory: memory)
+chain = LLMChain::Chain.new(model: "qwen3:1.7b", memory: memory, retriever: false)
+# retriever: false is required when you don't use a vector database to store context or external data
+# reitriever: - is set to WeaviateRetriever.new as default  so you need to pass an external params to set Weaviate host
 puts chain.ask("What is 2+2?")
 ```
 

data/lib/llm_chain/chain.rb

@@ -13,7 +13,13 @@ module LLMChain
       @model = model
       @memory = memory || Memory::Array.new
       @tools = tools
-      @retriever = retriever || Embeddings::Clients::Local::WeaviateRetriever.new
+      @retriever = if retriever.nil?
+                    Embeddings::Clients::Local::WeaviateRetriever.new
+                  elsif retriever == false
+                    nil
+                  else
+                    retriever
+                  end
       @client = ClientRegistry.client_for(model, **client_options)
     end
 
@@ -24,27 +30,20 @@ module LLMChain
     # @param rag_options [Hash] Опции для RAG-поиска
     # @yield [String] Передает чанки ответа если stream=true
     def ask(prompt, stream: false, rag_context: false, rag_options: {}, &block)
-      # 1. Сбор контекста
-      context = memory.recall(prompt)
-      tool_responses = process_tools(prompt)
-      rag_documents = retrieve_rag_context(prompt, rag_options) if rag_context
-
-      # 2. Построение промпта
-      full_prompt = build_prompt(
-        prompt: prompt,
-        memory_context: context,
-        tool_responses: tool_responses,
-        rag_documents: rag_documents
-      )
-
-      # 3. Генерация ответа
+      context = collect_context(prompt, rag_context, rag_options)
+      full_prompt = build_prompt(prompt: prompt, **context)
       response = generate_response(full_prompt, stream: stream, &block)
-
-      # 4. Сохранение в память
       memory.store(prompt, response)
       response
     end
 
+    def collect_context(prompt, rag_context, rag_options)
+      context = memory.recall(prompt)
+      tool_responses = process_tools(prompt)
+      rag_documents = retrieve_rag_context(prompt, rag_options) if rag_context
+      { memory_context: context, tool_responses: tool_responses, rag_documents: rag_documents }
+    end
+
     private
 
     def retrieve_rag_context(query, options = {})
@@ -53,8 +52,7 @@ module LLMChain
       limit = options[:limit] || 3
       @retriever.search(query, limit: limit)
     rescue => e
-      puts "[RAG Error] #{e.message}"
-      []
+      raise Error, "Cannot retrieve rag context"
     end
 
     def process_tools(prompt)
@@ -68,33 +66,37 @@ module LLMChain
 
     def build_prompt(prompt:, memory_context: nil, tool_responses: {}, rag_documents: nil)
       parts = []
+      parts << build_memory_context(memory_context) if memory_context&.any?
+      parts << build_rag_documents(rag_documents) if rag_documents&.any?
+      parts << build_tool_responses(tool_responses) unless tool_responses.empty?
+      parts << "Сurrent question: #{prompt}"
+      parts.join("\n\n")
+    end
 
-      if memory_context&.any?
-        parts << "Dialogue history:"
-        memory_context.each do |item|
-          parts << "User: #{item[:prompt]}"
-          parts << "Assistant: #{item[:response]}"
-        end
+    def build_memory_context(memory_context)
+      parts = ["Dialogue history:"]
+      memory_context.each do |item|
+        parts << "User: #{item[:prompt]}"
+        parts << "Assistant: #{item[:response]}"
       end
+      parts.join("\n")
+    end
 
-      if rag_documents&.any?
-        parts << "Relevant documents:"
-        rag_documents.each_with_index do |doc, i|
-          parts << "Document #{i + 1}: #{doc['content']}"
-          parts << "Metadata: #{doc['metadata'].to_json}" if doc['metadata']
-        end
+    def build_rag_documents(rag_documents)
+      parts = ["Relevant documents:"]
+      rag_documents.each_with_index do |doc, i|
+        parts << "Document #{i + 1}: #{doc['content']}"
+        parts << "Metadata: #{doc['metadata'].to_json}" if doc['metadata']
       end
+      parts.join("\n")
+    end
 
-      unless tool_responses.empty?
-        parts << "Tool results:"
-        tool_responses.each do |name, response|
-          parts << "#{name}: #{response}"
-        end
+    def build_tool_responses(tool_responses)
+      parts = ["Tool results:"]
+      tool_responses.each do |name, response|
+        parts << "#{name}: #{response}"
       end
-
-      parts << "Qurrent question: #{prompt}"
-
-      parts.join("\n\n")
+      parts.join("\n")
     end
 
     def generate_response(prompt, stream: false, &block)

data/lib/llm_chain/client_registry.rb

@@ -15,6 +15,8 @@ module LLMChain
         Clients::Qwen
       when /llama2/
         Clients::Llama2
+      when /gemma3/
+        Clients::Gemma3
       else
         raise UnknownModelError, "Unknown model: #{model}"
       end

data/lib/llm_chain/clients/gemma3.rb

@@ -0,0 +1,144 @@
+require 'faraday'
+require 'json'
+
+module LLMChain
+  module Clients
+    class Gemma3 < OllamaBase
+      # Доступные версии моделей Gemma3
+      MODEL_VERSIONS = {
+        gemma3: {
+          default: "gemma3:2b",
+          versions: [
+            "gemma3:2b", "gemma3:8b", "gemma3:27b",
+            "gemma3:2b-instruct", "gemma3:8b-instruct", "gemma3:27b-instruct", "gemma3:4b"
+          ]
+        }
+      }.freeze
+
+      # Общие настройки по умолчанию для Gemma3
+      COMMON_DEFAULT_OPTIONS = {
+        temperature: 0.7,
+        top_p: 0.9,
+        top_k: 40,
+        repeat_penalty: 1.1,
+        num_ctx: 8192
+      }.freeze
+
+      # Специфичные настройки для разных версий
+      VERSION_SPECIFIC_OPTIONS = {
+        gemma3: {
+          stop: ["<|im_end|>", "<|endoftext|>", "<|user|>", "<|assistant|>"]
+        }
+      }.freeze
+
+      # Внутренние теги для очистки ответов
+      INTERNAL_TAGS = {
+        common: {
+          think: /<think>.*?<\/think>\s*/mi,
+          reasoning: /<reasoning>.*?<\/reasoning>\s*/mi
+        },
+        gemma3: {
+          system: /<\|system\|>.*?<\|im_end\|>\s*/mi,
+          user: /<\|user\|>.*?<\|im_end\|>\s*/mi,
+          assistant: /<\|assistant\|>.*?<\|im_end\|>\s*/mi
+        }
+      }.freeze
+
+      def initialize(model: nil, base_url: nil, **options)
+        model ||= detect_default_model
+
+        @model = model
+        validate_model_version(@model)
+
+        super(
+          model: @model,
+          base_url: base_url,
+          default_options: default_options_for(@model).merge(options)
+        )
+      end
+
+      def chat(prompt, show_internal: false, stream: false, **options, &block)
+        if stream
+          stream_chat(prompt, show_internal: show_internal, **options, &block)
+        else
+          response = super(prompt, **options)
+          process_response(response, show_internal: show_internal)
+        end
+      end
+
+      def stream_chat(prompt, show_internal: false, **options, &block)
+        buffer = ""
+        connection.post(API_ENDPOINT) do |req|
+          req.headers['Content-Type'] = 'application/json'
+          req.body = build_request_body(prompt, options.merge(stream: true))
+          
+          req.options.on_data = Proc.new do |chunk, _bytes, _env|
+            processed = process_stream_chunk(chunk)
+            next unless processed
+
+            buffer << processed
+            block.call(processed) if block_given?
+          end
+        end
+
+        process_response(buffer, show_internal: show_internal)
+      end
+
+      protected
+
+      def build_request_body(prompt, options)
+        body = super
+        version_specific_options = VERSION_SPECIFIC_OPTIONS[model_version]
+        body[:options].merge!(version_specific_options) if version_specific_options
+        body
+      end
+
+      private
+
+      def model_version
+        :gemma3
+      end
+
+      def detect_default_model
+        MODEL_VERSIONS[model_version][:default]
+      end
+
+      def validate_model_version(model)
+        valid_models = MODEL_VERSIONS.values.flat_map { |v| v[:versions] }
+        unless valid_models.include?(model)
+          raise InvalidModelVersion, "Invalid Gemma3 model version. Available: #{valid_models.join(', ')}"
+        end
+      end
+
+      def default_options_for(model)
+        COMMON_DEFAULT_OPTIONS.merge(
+          VERSION_SPECIFIC_OPTIONS[model_version] || {}
+        )
+      end
+
+      def process_stream_chunk(chunk)
+        parsed = JSON.parse(chunk)
+        parsed["response"] if parsed.is_a?(Hash) && parsed["response"]
+      rescue JSON::ParserError
+        nil
+      end
+
+      def process_response(response, show_internal: false)
+        return response unless response.is_a?(String)
+        
+        if show_internal
+          response
+        else
+          clean_response(response)
+        end
+      end
+
+      def clean_response(text)
+        tags = INTERNAL_TAGS[:common].merge(INTERNAL_TAGS[model_version] || {})
+        tags.values.reduce(text) do |processed, regex|
+          processed.gsub(regex, '')
+        end.gsub(/\n{3,}/, "\n\n").strip
+      end
+    end
+  end
+end 

data/lib/llm_chain/clients/qwen.rb

@@ -92,9 +92,8 @@ module LLMChain
 
       def build_request_body(prompt, options)
         body = super
-        version_specific_options = VERSION_SPECIFIC_OPTIONS[model_version]
+        version_specific_options = VERSION_SPECIFIC_OPTIONS[model_version]  
         body[:options].merge!(version_specific_options) if version_specific_options
-        puts body
         body
       end
 

data/lib/llm_chain/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module LlmChain
-  VERSION = "0.3.0"
+  VERSION = "0.4.0"
 end

data/lib/llm_chain.rb

@@ -17,6 +17,7 @@ require "llm_chain/clients/ollama_base"
 require "llm_chain/clients/openai"
 require "llm_chain/clients/qwen"
 require "llm_chain/clients/llama2"
+require "llm_chain/clients/gemma3"
 require "llm_chain/client_registry"
 require "llm_chain/memory/array"
 require "llm_chain/memory/redis"

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: llm_chain
 version: !ruby/object:Gem::Version
-  version: 0.3.0
+  version: 0.4.0
 platform: ruby
 authors:
 - FuryCow
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2025-06-16 00:00:00.000000000 Z
+date: 2025-06-23 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: httparty
@@ -125,6 +125,7 @@ files:
 - lib/llm_chain/chain.rb
 - lib/llm_chain/client_registry.rb
 - lib/llm_chain/clients/base.rb
+- lib/llm_chain/clients/gemma3.rb
 - lib/llm_chain/clients/llama2.rb
 - lib/llm_chain/clients/ollama_base.rb
 - lib/llm_chain/clients/openai.rb
@@ -160,5 +161,5 @@ requirements: []
 rubygems_version: 3.4.10
 signing_key:
 specification_version: 4
-summary: Ruby-аналог LangChain для работы с LLM
+summary: Ruby-analog LangChain to work with LLM
 test_files: []