legion-llm 0.5.21 → 0.5.22

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: '08ccdbe9c1f4187acdcae49df633370d7fecb9866fa441aa1e95a3f932d5c9e6'
-  data.tar.gz: df48f1ba0ef83a0fb26ba67865bf0f69ee415a349f2fc0e4ec40dd70f7f70815
+  metadata.gz: 7e7ff9c35609294ace1f203076485aaba0a811ba0adbe7d01197ce91de7a06ea
+  data.tar.gz: c493d891ded32632daf9b7e367fa115efc9903cc63d408c8c0799bb8bc242ca1
 SHA512:
-  metadata.gz: 9237a7a67d3b843bef628817cbd679fdce2f690e35b35f817e38b39c2e46918cb87d1c81b0b3c5a48589f54a163cbd7f3ef5956e73ba3c96f4f0ee3c1f803c9d
-  data.tar.gz: 996f35c9bb47ff5046bfd10fe48440a783480b81956b6e0b7f1f10bd860fd4aaab9dd9fcda3378761ac2942ea7f9c886d0bd19722cee9b46936f7ae8c662aa9d
+  metadata.gz: 2cd3766d938ba6814f2a14af190441a0d9b0e011538da22744729b678313ff03045fc84ae4e1ecab7dcff9ac31a1d4d92afdff25d3b5e7ca322a0303ef54c0fa
+  data.tar.gz: 3a03dd436e31b0d02290e23af4cadc755136c89321af205a83c0c18dac4cac9a8f320875a578ac6405dafd409a0b3bc516431d1dff34670731ab803618df36ae

data/CHANGELOG.md

@@ -2,6 +2,18 @@
 
 ## [Unreleased]
 
+## [0.5.22] - 2026-03-31
+
+### Added
+- Auto-chunking for oversized Ollama embedding inputs via `lex-knowledge` Chunker with character-split fallback
+- `average_vectors` for document-level embedding from multiple chunks
+- Per-model Ollama context limits (`OLLAMA_CONTEXT_CHARS`): mxbai-embed-large 2048, nomic-embed-text 32768
+- `lex-knowledge` added as a dependency for semantic chunking
+
+### Fixed
+- `handle_embed_failure` no longer permanently mutates `@embedding_provider` — failover is per-request only
+- `ollama_preferred` order corrected: `mxbai-embed-large` (1024 dims) first, `nomic-embed-text` (768 dims) second
+
 ## [0.5.21] - 2026-03-31
 
 ### Added

data/legion-llm.gemspec

@@ -33,6 +33,7 @@ Gem::Specification.new do |spec|
   spec.add_dependency 'lex-bedrock'
   spec.add_dependency 'lex-claude'
   spec.add_dependency 'lex-gemini'
+  spec.add_dependency 'lex-knowledge'
   spec.add_dependency 'lex-openai'
   spec.add_dependency 'ruby_llm', '~> 1.13'
   spec.add_dependency 'tzinfo', '>= 2.0'

data/lib/legion/llm/embeddings.rb

@@ -16,6 +16,14 @@ module Legion
 
       TARGET_DIMENSION = 1024
 
+      OLLAMA_CONTEXT_CHARS = {
+        'mxbai-embed-large'      => 2048,
+        'bge-large'              => 2048,
+        'snowflake-arctic-embed' => 2048,
+        'nomic-embed-text'       => 32_768
+      }.freeze
+      OLLAMA_DEFAULT_CONTEXT_CHARS = 2048
+
       class << self
         def generate(text:, model: nil, provider: nil, dimensions: nil)
           return { vector: nil, model: model, provider: provider, error: 'LLM not started' } unless LLM.started?
@@ -97,8 +105,6 @@ module Legion
           fallback = find_fallback_provider(failed_provider)
           if fallback
             Legion::Logging.info "Embedding failover: #{failed_provider} -> #{fallback[:provider]}" if defined?(Legion::Logging)
-            LLM.instance_variable_set(:@embedding_provider, fallback[:provider])
-            LLM.instance_variable_set(:@embedding_model, fallback[:model])
             generate(text: text, model: fallback[:model], provider: fallback[:provider])
           else
             { vector: nil, model: failed_model, provider: failed_provider, error: error.message }
@@ -177,6 +183,9 @@ module Legion
         end
 
         def generate_ollama(text:, model:)
+          max_chars = ollama_context_chars(model)
+          return generate_ollama_chunked(text: text, model: model, max_chars: max_chars) if text.length > max_chars
+
           result = ollama_embed_request(model: model, input: text)
           vector = result['embeddings']&.first
           vector = apply_dimension_enforcement(vector, :ollama) if vector
@@ -185,14 +194,63 @@ module Legion
           { vector: vector, model: model, provider: :ollama, dimensions: vector&.size || 0, tokens: 0 }
         end
 
+        def generate_ollama_chunked(text:, model:, max_chars:)
+          chunks = chunk_text(text, max_chars: max_chars)
+          vectors = chunks.filter_map do |chunk|
+            result = ollama_embed_request(model: model, input: chunk[:content])
+            result['embeddings']&.first
+          end
+
+          return { vector: nil, model: model, provider: :ollama, error: 'all chunks failed embedding' } if vectors.empty?
+
+          avg = average_vectors(vectors)
+          avg = apply_dimension_enforcement(avg, :ollama)
+          return dimension_error(model, :ollama, avg) if avg.is_a?(String)
+
+          { vector: avg, model: model, provider: :ollama, dimensions: avg.size, tokens: 0, chunks: vectors.size }
+        end
+
         def generate_ollama_batch(texts:, model:)
-          result = ollama_embed_request(model: model, input: texts)
-          vectors = result['embeddings'] || []
-          vectors.each_with_index.map do |vec, i|
-            build_batch_entry(vec, model, :ollama, i)
+          max_chars = ollama_context_chars(model)
+          texts.each_with_index.map do |text, i|
+            if text.length > max_chars
+              result = generate_ollama_chunked(text: text, model: model, max_chars: max_chars)
+              build_batch_entry(result[:vector], model, :ollama, i)
+            else
+              result = ollama_embed_request(model: model, input: text)
+              vec = result['embeddings']&.first
+              build_batch_entry(vec, model, :ollama, i)
+            end
           end
         end
 
+        def chunk_text(text, max_chars:)
+          if defined?(Legion::Extensions::Knowledge::Helpers::Chunker)
+            chunker = Legion::Extensions::Knowledge::Helpers::Chunker
+            max_tokens = max_chars / chunker::CHARS_PER_TOKEN
+            sections = [{ content: text, heading: nil, section_path: nil, source_file: nil }]
+            chunker.chunk(sections: sections, max_tokens: max_tokens)
+          else
+            text.chars.each_slice(max_chars).map { |s| { content: s.join } }
+          end
+        rescue StandardError
+          text.chars.each_slice(max_chars).map { |s| { content: s.join } }
+        end
+
+        def average_vectors(vectors)
+          return vectors.first if vectors.size == 1
+
+          dim = vectors.first.size
+          sum = Array.new(dim, 0.0)
+          vectors.each { |v| v.each_with_index { |val, i| sum[i] += val } }
+          sum.map { |s| s / vectors.size }
+        end
+
+        def ollama_context_chars(model)
+          base = model.to_s.split(':').first
+          OLLAMA_CONTEXT_CHARS[base] || OLLAMA_DEFAULT_CONTEXT_CHARS
+        end
+
         def ollama_embed_request(model:, input:)
           base_url = Legion::Settings.dig(:llm, :providers, :ollama, :base_url) || 'http://localhost:11434'
           conn = Faraday.new(url: base_url) do |f|

data/lib/legion/llm/settings.rb

@@ -150,7 +150,7 @@ module Legion
             bedrock: 'amazon.titan-embed-text-v2:0',
             openai:  'text-embedding-3-small'
           },
-          ollama_preferred:  %w[nomic-embed-text mxbai-embed-large bge-large snowflake-arctic-embed]
+          ollama_preferred:  %w[mxbai-embed-large nomic-embed-text bge-large snowflake-arctic-embed]
         }
       end
 

data/lib/legion/llm/version.rb

@@ -2,6 +2,6 @@
 
 module Legion
   module LLM
-    VERSION = '0.5.21'
+    VERSION = '0.5.22'
   end
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: legion-llm
 version: !ruby/object:Gem::Version
-  version: 0.5.21
+  version: 0.5.22
 platform: ruby
 authors:
 - Esity
@@ -121,6 +121,20 @@ dependencies:
     - - ">="
       - !ruby/object:Gem::Version
         version: '0'
+- !ruby/object:Gem::Dependency
+  name: lex-knowledge
+  requirement: !ruby/object:Gem::Requirement
+    requirements:
+    - - ">="
+      - !ruby/object:Gem::Version
+        version: '0'
+  type: :runtime
+  prerelease: false
+  version_requirements: !ruby/object:Gem::Requirement
+    requirements:
+    - - ">="
+      - !ruby/object:Gem::Version
+        version: '0'
 - !ruby/object:Gem::Dependency
   name: lex-openai
   requirement: !ruby/object:Gem::Requirement