ruby_llm 2.0.0.rc1 → 2.0.0.rc3

data/lib/ruby_llm/protocols/anthropic/chat.rb

@@ -18,6 +18,7 @@ module RubyLLM
         COMPACTION_BETA = 'compact-2026-01-12'
         COMPACTION_EDIT_TYPE = 'compact_20260112'
         COUNT_TOKENS_KEYS = %i[model messages system tools tool_choice thinking].freeze
+        THINKING_BLOCK_TYPES = %w[thinking redacted_thinking].freeze
 
         module_function
 
@@ -303,6 +304,11 @@ module RubyLLM
           thinking_block&.dig('signature') || thinking_block&.dig('data')
         end
 
+        def parse_thinking_blocks(blocks)
+          thinking = blocks.select { |block| THINKING_BLOCK_TYPES.include?(block['type']) }
+          { 'anthropic' => thinking } unless thinking.empty?
+        end
+
         def build_message(data, content:, citations:, thinking:, thinking_signature:, tool_use_blocks:, raw:,
                           server_tool_calls: [], raw_content: nil)
           usage = aggregate_usage(data['usage'])
@@ -316,6 +322,7 @@ module RubyLLM
             content: content,
             citations: citations,
             thinking: Thinking.build(text: thinking, signature: thinking_signature),
+            raw_reasoning: parse_thinking_blocks(data['content'] || []),
             tool_calls: Tools.parse_tool_calls(tool_use_blocks),
             server_tool_calls: server_tool_calls,
             raw_content: raw_content,
@@ -357,12 +364,7 @@ module RubyLLM
         end
 
         def format_basic_message_with_thinking(msg, citations: false, caching: nil)
-          content_blocks = []
-
-          if msg.role == :assistant
-            thinking_block = build_thinking_block(msg.thinking)
-            content_blocks << thinking_block if thinking_block
-          end
+          content_blocks = msg.role == :assistant ? format_thinking_blocks(msg) : []
 
           append_formatted_content(content_blocks, msg, citations: citations)
           inject_cache_control(content_blocks, caching:) if cache_boundary?(msg, caching:)
@@ -374,7 +376,7 @@ module RubyLLM
         end
 
         def format_tool_call_with_thinking(msg, caching: nil)
-          content_blocks = prepend_thinking_block([], msg)
+          content_blocks = prepend_thinking_blocks([], msg)
           append_formatted_content(content_blocks, msg) unless msg.content.nil? || msg.content.empty?
 
           msg.tool_calls.each_value do |tool_call|
@@ -393,13 +395,19 @@ module RubyLLM
           }
         end
 
-        def prepend_thinking_block(content_blocks, msg)
-          thinking_block = build_thinking_block(msg.thinking)
-          content_blocks.unshift(thinking_block) if thinking_block
+        def prepend_thinking_blocks(content_blocks, msg)
+          content_blocks.unshift(*format_thinking_blocks(msg))
 
           content_blocks
         end
 
+        def format_thinking_blocks(msg)
+          blocks = msg.raw_reasoning['anthropic'] if msg.raw_reasoning.is_a?(Hash)
+          return Support::Utils.deep_dup(blocks) if blocks
+
+          [build_thinking_block(msg.thinking)].compact
+        end
+
         def build_thinking_block(thinking)
           return nil unless thinking
 

data/lib/ruby_llm/protocols/anthropic/streaming.rb

@@ -108,12 +108,13 @@ module RubyLLM
         end
 
         def stream_end_fields(data)
-          return {} unless data['type'] == 'message_stop' && @saw_server_block
+          return {} unless data['type'] == 'message_stop'
 
-          blocks = @stream_blocks.sort.map { |_index, block| block }
+          blocks = (@stream_blocks || {}).sort.map { |_index, block| block }
           {
             server_tool_calls: extract_server_tool_calls(blocks),
-            raw_content: blocks
+            raw_content: @saw_server_block ? blocks : nil,
+            raw_reasoning: parse_thinking_blocks(blocks)
           }
         end
 

data/lib/ruby_llm/protocols/anthropic.rb

@@ -87,8 +87,9 @@ module RubyLLM
       end
 
       def merge_raw_content(segments)
-        merged = segments.flat_map { |segment| segment.raw_content || [] }
-        merged.empty? ? nil : merged
+        return unless segments.any?(&:raw_content)
+
+        segments.flat_map { |segment| segment.raw_content || format_message(segment)[:content] }
       end
 
       def merge_tool_calls(segments)

data/lib/ruby_llm/protocols/bedrock/async_videos.rb

@@ -64,7 +64,8 @@ module RubyLLM
           prefix = job.raw.dig('outputDataConfig', 's3OutputDataConfig', 's3Uri')
           raise Error, 'Bedrock video job has no output S3 URI' unless prefix
 
-          uris = @provider.list_file_uris("#{prefix.sub(%r{/+\z}, '')}/")
+          prefix_length = (prefix.rindex(%r{[^/]}) || -1) + 1
+          uris = @provider.list_file_uris("#{prefix[0, prefix_length]}/")
           videos = uris.select { |uri| uri.downcase.end_with?('.mp4') }
           raise Error, 'Expected exactly one MP4 in the Bedrock video output' unless videos.one?
 

data/lib/ruby_llm/protocols/chat_completions/rerank.rb

@@ -38,10 +38,13 @@ module RubyLLM
         def parse_rerank_results(data, documents = [])
           Array(data['results']).map do |result|
             index = result['index']
+            unless valid_rerank_index?(index, documents)
+              raise Error, 'Rerank endpoint returned an invalid document index'
+            end
 
             RubyLLM::Rerank::Result.new(
               index: index,
-              document: rerank_document(result['document']) || (index && documents[index]),
+              document: rerank_document(result['document']) || documents[index],
               score: result['relevance_score']
             )
           end
@@ -50,6 +53,10 @@ module RubyLLM
         def rerank_document(document)
           document.is_a?(Hash) ? document['text'] : document
         end
+
+        def valid_rerank_index?(index, documents)
+          index.is_a?(Integer) && index.between?(0, documents.length - 1)
+        end
       end
     end
   end

data/lib/ruby_llm/protocols/cohere/rerank.rb

@@ -38,14 +38,21 @@ module RubyLLM
         def parse_rerank_results(data, documents)
           Array(data['results']).map do |result|
             index = result['index']
+            unless valid_rerank_index?(index, documents)
+              raise Error, 'Cohere reranking returned an invalid document index'
+            end
 
             RubyLLM::Rerank::Result.new(
               index: index,
-              document: result.dig('document', 'text') || (index && documents[index]),
+              document: result.dig('document', 'text') || documents[index],
               score: result['relevance_score']
             )
           end
         end
+
+        def valid_rerank_index?(index, documents)
+          index.is_a?(Integer) && index.between?(0, documents.length - 1)
+        end
       end
     end
   end

data/lib/ruby_llm/protocols/converse/chat.rb

@@ -117,6 +117,7 @@ module RubyLLM
             content: text_content,
             citations: citations,
             thinking: Thinking.build(text: thinking_text, signature: thinking_signature),
+            raw_reasoning: parse_thinking_blocks(content_blocks),
             tool_calls: parse_tool_calls(content_blocks),
             server_tool_calls: extract_server_tool_calls(content_blocks),
             input_tokens: input_tokens(usage),
@@ -195,10 +196,7 @@ module RubyLLM
         end
 
         def format_structured_message_content(msg, citations: false)
-          blocks = []
-
-          thinking_block = format_thinking_block(msg.thinking)
-          blocks << thinking_block if msg.role == :assistant && thinking_block
+          blocks = msg.role == :assistant ? format_thinking_blocks(msg) : []
 
           blocks.concat(
             Media.format_content(msg.content, msg.attachments, used_document_names: @used_document_names, citations:)
@@ -207,6 +205,13 @@ module RubyLLM
           blocks
         end
 
+        def format_thinking_blocks(msg)
+          blocks = msg.raw_reasoning['converse'] if msg.raw_reasoning.is_a?(Hash)
+          return Support::Utils.deep_dup(blocks) if blocks
+
+          [format_thinking_block(msg.thinking)].compact
+        end
+
         def format_tool_result_block(msg)
           {
             toolResult: {
@@ -614,6 +619,11 @@ module RubyLLM
           [text, signature]
         end
 
+        def parse_thinking_blocks(content_blocks)
+          blocks = content_blocks.select { |block| block['reasoningContent'].is_a?(Hash) }
+          { 'converse' => blocks } unless blocks.empty?
+        end
+
         def parse_reasoning_content_block(block)
           reasoning_content = block['reasoningContent']
           return [nil, nil] unless reasoning_content.is_a?(Hash)

data/lib/ruby_llm/protocols/converse/streaming.rb

@@ -18,6 +18,7 @@ module RubyLLM
         end
 
         def stream_response(payload, additional_headers = {}, &block)
+          @thinking_stream = ThinkingStream.new
           accumulator = RubyLLM::Protocol::StreamAccumulator.new
           decoder = event_stream_decoder
           body = JSON.generate(payload)
@@ -195,6 +196,7 @@ module RubyLLM
               text: extract_thinking_delta(event),
               signature: extract_thinking_signature(event)
             ),
+            raw_reasoning: streamed_thinking_blocks(event),
             tool_calls: extract_tool_calls(event),
             server_tool_calls: extract_server_tool_call_events(event),
             input_tokens: extract_input_tokens(metadata_usage, usage),
@@ -206,6 +208,12 @@ module RubyLLM
           )
         end
 
+        def streamed_thinking_blocks(event)
+          @thinking_stream ||= ThinkingStream.new
+          @thinking_stream.add(event)
+          @thinking_stream.raw_reasoning if event.key?('messageStop') || event.key?('stopReason')
+        end
+
         def extract_finish_reason(event)
           normalize_finish_reason(event.dig('messageStop', 'stopReason') || event['stopReason'])
         end

data/lib/ruby_llm/protocols/converse/thinking_stream.rb

@@ -0,0 +1,56 @@
+# frozen_string_literal: true
+
+require 'base64'
+
+module RubyLLM
+  module Protocols
+    class Converse
+      class ThinkingStream # :nodoc: all
+        def initialize
+          @blocks = {}
+        end
+
+        def add(event)
+          @blocks.clear if event.key?('messageStart')
+          part = event['contentBlockStart'] || event['contentBlockDelta'] || event
+          index = part['contentBlockIndex']
+          return if index.nil?
+
+          start = part.dig('start', 'reasoningContent')
+          @blocks[index] = Support::Utils.deep_dup(start) if start
+          delta = part.dig('delta', 'reasoningContent')
+          append_delta(index, delta) if delta
+        end
+
+        def raw_reasoning
+          return if @blocks.empty?
+
+          { 'converse' => @blocks.sort.map { |_index, content| { 'reasoningContent' => content } } }
+        end
+
+        private
+
+        def append_delta(index, delta)
+          block = (@blocks[index] ||= {})
+          if delta.key?('redactedContent')
+            append_redacted(block, delta['redactedContent'])
+          else
+            text = (block['reasoningText'] ||= { 'text' => '' })
+            (delta['reasoningText'] || delta).slice('text', 'signature').each do |key, value|
+              text[key] = text[key].to_s + value.to_s
+            end
+          end
+        end
+
+        def append_redacted(block, data)
+          block['redactedContent'] = if block['redactedContent']
+                                       Base64.strict_encode64(Base64.decode64(block['redactedContent']) +
+                                                              Base64.decode64(data))
+                                     else
+                                       data
+                                     end
+        end
+      end
+    end
+  end
+end

data/lib/ruby_llm/protocols/gemini/embedding_batches.rb

@@ -60,6 +60,11 @@ module RubyLLM
           metadata = responses.first.fetch('metadata')
           return if responses.size < metadata.fetch('embedding_count')
 
+          positions = responses.map { |inline| inline.dig('metadata', 'embedding_index') }
+          unless positions.sort == (0...responses.size).to_a
+            return [index, nil, batch_failure(key, 'Invalid or duplicate embedding record positions')]
+          end
+
           vectors = embedding_batch_vectors(responses)
           return [index, nil, batch_failure(key, 'Gemini returned no embedding')] unless vectors
 

data/lib/ruby_llm/protocols/interactions/tools.rb

@@ -31,7 +31,9 @@ module RubyLLM
         end
 
         def parse_interaction_arguments(arguments)
-          arguments.is_a?(String) ? JSON.parse(arguments) : arguments || {}
+          return {} if arguments.nil? || (arguments.is_a?(String) && arguments.empty?)
+
+          arguments.is_a?(String) ? JSON.parse(arguments) : arguments
         rescue JSON::ParserError => e
           raise ToolCallParseError.new(finish_reason: :tool_calls), cause: e
         end

data/lib/ruby_llm/providers/deepseek/responses.rb

@@ -8,7 +8,6 @@ module RubyLLM
       # instead of summaries, both in responses and in the stream.
       class Responses < Protocols::Responses
         SERVER_TOOL_ALIASES = {
-          web_search: { tool: { type: 'web_search' } },
           apply_patch: { tool: { type: 'custom', name: 'apply_patch' } }
         }.freeze
 

data/lib/ruby_llm/providers/mistral/ocr.rb

@@ -22,7 +22,7 @@ module RubyLLM
         end
 
         def ocr_document_part(attachment)
-          reference = attachment.url? ? attachment.source.to_s : attachment.for_llm
+          reference = attachment.url? ? attachment.source.to_s : ocr_data_uri(attachment)
 
           if attachment.image?
             { type: 'image_url', image_url: reference }
@@ -31,6 +31,10 @@ module RubyLLM
           end
         end
 
+        def ocr_data_uri(attachment)
+          "data:#{attachment.mime_type};base64,#{attachment.encoded}"
+        end
+
         def parse_ocr_response(response, model:)
           data = response.body
 

data/lib/ruby_llm/providers/openrouter/chat.rb

@@ -80,7 +80,7 @@ module RubyLLM
 
         def format_thinking(msg)
           return {} unless msg.role == :assistant
-          return { reasoning_details: msg.raw_reasoning } if msg.raw_reasoning
+          return { reasoning_details: msg.raw_reasoning } if msg.raw_reasoning.is_a?(Array)
 
           thinking = msg.thinking
           return {} unless thinking

data/lib/ruby_llm/version.rb

@@ -2,5 +2,5 @@
 
 module RubyLLM
   # The version of the ruby_llm gem, as a string.
-  VERSION = '2.0.0.rc1'
+  VERSION = '2.0.0.rc3'
 end

data/lib/tasks/ruby_llm.rake

@@ -7,7 +7,7 @@ namespace :ruby_llm do
       task action => :environment do
         require 'generators/ruby_llm/upgrade/upgrade_migration'
 
-        RubyLLM::Generators::UpgradeMigration.new.public_send(action)
+        RubyLLM::Generators::UpgradeMigration.for.public_send(action)
         puts "RubyLLM copy upgrade: #{action} completed"
       end
     end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: ruby_llm
 version: !ruby/object:Gem::Version
-  version: 2.0.0.rc1
+  version: 2.0.0.rc3
 platform: ruby
 authors:
 - Carmine Paolino
@@ -275,6 +275,11 @@ files:
 - lib/generators/ruby_llm/tool/templates/tool_call.html.erb.tt
 - lib/generators/ruby_llm/tool/templates/tool_result.html.erb.tt
 - lib/generators/ruby_llm/tool/tool_generator.rb
+- lib/generators/ruby_llm/upgrade/legacy_content_sql.rb
+- lib/generators/ruby_llm/upgrade/online_copy_migration.rb
+- lib/generators/ruby_llm/upgrade/online_copy_migration/data.rb
+- lib/generators/ruby_llm/upgrade/online_copy_migration/journal.rb
+- lib/generators/ruby_llm/upgrade/online_copy_migration/verification.rb
 - lib/generators/ruby_llm/upgrade/templates/backfill_v2_data.rb.tt
 - lib/generators/ruby_llm/upgrade/templates/cleanup_v2_upgrade.rb.tt
 - lib/generators/ruby_llm/upgrade/templates/finish_v2_upgrade.rb.tt
@@ -379,6 +384,7 @@ files:
 - lib/ruby_llm/protocols/converse/chat.rb
 - lib/ruby_llm/protocols/converse/media.rb
 - lib/ruby_llm/protocols/converse/streaming.rb
+- lib/ruby_llm/protocols/converse/thinking_stream.rb
 - lib/ruby_llm/protocols/deepgram.rb
 - lib/ruby_llm/protocols/deepgram/models.rb
 - lib/ruby_llm/protocols/deepgram/speech.rb
@@ -611,7 +617,7 @@ metadata:
   funding_uri: https://github.com/sponsors/crmne
   rubygems_mfa_required: 'true'
 post_install_message: |
-  RubyLLM 2.0 release candidate 1
+  RubyLLM 2.0.0.rc3
 
     2.0 renames several APIs and changes what message content returns. Coming
     from 1.x? Read the upgrade guide before you boot: