legion-llm 0.6.0 → 0.6.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: f65ac724c32de98ddfa324545b62e81cda38e27efdcdcbceb21abd21729ae599
-  data.tar.gz: 7c02a90eac3bda99512da956c889a06084980468c034c25e0c602d7e06db7ac3
+  metadata.gz: f52899e131c7b40d4dba1576584fa0e9709022db8c83dfc6c2bf21044c58efe4
+  data.tar.gz: a45b7ba54bd018efbdaa61c1b13d53608560c432cb804ccb565aea5e99e0087e
 SHA512:
-  metadata.gz: 71f7496e4df651c8d93bf3ac27059a2075f0b82299afa1f61f98138dc81db90ed3139c27b933774969d7d727ff9483db2a92514d82460a3f3de1c2dfbbff44ff
-  data.tar.gz: 6757e931ab1bef7d95c1470a3cf24077fa777683955bc0e5ed6ab6b7d7ef6a2f6f4f613668b6395a52a4200cdb97ae9aebc80f6bc9af5ee896c8a44142215425
+  metadata.gz: ea38c1ca6a6298e7aae10524afa822a2e85a83d0958ea8c830cf046470a3260db3605a0a59fed85bca08939a1914c86e226e306a11bbb602ffb6b6b13cfa6181
+  data.tar.gz: a57924ff5d697c34160d1fb5aa4ce515f55a7e0f52e5bab8d740bcbe31bfb6a947d475e44404f3f84b8078441df6087f1b8f466d25a0eb6e937781c601c99b7b

data/CHANGELOG.md

@@ -2,6 +2,19 @@
 
 ## [Unreleased]
 
+## [0.6.2] - 2026-03-31
+
+### Fixed
+- Reduce `OLLAMA_CONTEXT_CHARS` from 2048 to 1400 for 512-token models (mxbai-embed-large, bge-large, snowflake-arctic-embed) to account for real tokenization ratios (~3 chars/token)
+- `generate_ollama` now catches context-length rejections and retries with chunking at 60% char limit instead of failing over to a potentially broken provider
+
+## [0.6.1] - 2026-03-31
+
+### Added
+- Advisory step reads calibration_weights from Apollo Local, includes in advisory enrichment
+- Advisory meta recording: classifies advisory types and calls `Legion::Gaia.record_advisory_meta`
+- Advisory type classification based on partner context (tone, verbosity, format, context, hint)
+
 ## [0.6.0] - 2026-03-31
 
 ### Added

data/lib/legion/llm/embeddings.rb

@@ -17,12 +17,12 @@ module Legion
       TARGET_DIMENSION = 1024
 
       OLLAMA_CONTEXT_CHARS = {
-        'mxbai-embed-large'      => 2048,
-        'bge-large'              => 2048,
-        'snowflake-arctic-embed' => 2048,
-        'nomic-embed-text'       => 32_768
+        'mxbai-embed-large'      => 1400,
+        'bge-large'              => 1400,
+        'snowflake-arctic-embed' => 1400,
+        'nomic-embed-text'       => 24_000
       }.freeze
-      OLLAMA_DEFAULT_CONTEXT_CHARS = 2048
+      OLLAMA_DEFAULT_CONTEXT_CHARS = 1400
 
       PREFIX_REGISTRY = {
         'nomic-embed-text'  => { document: 'search_document: ', query: 'search_query: ' },
@@ -219,6 +219,12 @@ module Legion
           return dimension_error(model, :ollama, vector) if vector.is_a?(String)
 
           { vector: vector, model: model, provider: :ollama, dimensions: vector&.size || 0, tokens: 0 }
+        rescue RuntimeError => e
+          raise unless e.message.include?('input length exceeds')
+
+          reduced = (max_chars * 0.6).to_i
+          Legion::Logging.info("Ollama context exceeded, retrying with chunking at #{reduced} chars") if defined?(Legion::Logging)
+          generate_ollama_chunked(text: text, model: model, max_chars: reduced)
         end
 
         def generate_ollama_chunked(text:, model:, max_chars:)

data/lib/legion/llm/pipeline/steps/gaia_advisory.rb

@@ -21,6 +21,9 @@ module Legion
 
             enrich_advisory_with_partner_context(advisory)
 
+            calibration_weights = fetch_calibration_weights
+            advisory[:calibration_weights] = calibration_weights if calibration_weights
+
             @enrichments['gaia:advisory'] = {
               content:   advisory_summary(advisory),
               data:      advisory,
@@ -46,6 +49,8 @@ module Legion
               direction: :inbound, detail: advisory_summary(advisory),
               from: 'gaia', to: 'pipeline'
             )
+
+            record_advisory_meta_to_gaia(advisory)
           rescue StandardError => e
             @warnings << "GAIA advisory error: #{e.message}"
           end
@@ -160,6 +165,49 @@ module Legion
               :infrequent
             end
           end
+
+          def fetch_calibration_weights
+            return nil unless apollo_local_available?
+
+            result = ::Legion::Apollo::Local.query(
+              text: 'bond calibration weights',
+              tags: %w[bond calibration weights]
+            )
+            return nil unless result[:success] && result[:results]&.any?
+
+            raw = ::JSON.parse(result[:results].first[:content])
+            raw['weights']
+          rescue StandardError
+            nil
+          end
+
+          def record_advisory_meta_to_gaia(advisory)
+            return unless defined?(::Legion::Gaia) && ::Legion::Gaia.respond_to?(:record_advisory_meta)
+            return unless advisory[:partner_context]
+
+            advisory_id = SecureRandom.uuid
+            advisory_types = classify_advisory_types(advisory)
+
+            ::Legion::Gaia.record_advisory_meta(
+              advisory_id:    advisory_id,
+              advisory_types: advisory_types
+            )
+          rescue StandardError
+            nil
+          end
+
+          def classify_advisory_types(advisory)
+            types = []
+            pc = advisory[:partner_context]
+            return ['partner_hint'] unless pc
+
+            types << 'partner_hint' if pc
+            types << 'context_injection' if advisory[:context_window]
+            types << 'tone_adjustment' if pc[:recent_sentiment] && pc[:recent_sentiment] != :neutral
+            types << 'verbosity_adjustment' if pc[:interaction_pattern] && pc[:interaction_pattern] != :unknown
+            types << 'format_adjustment' if pc[:compatibility]
+            types.empty? ? ['partner_hint'] : types
+          end
         end
       end
     end

data/lib/legion/llm/version.rb

@@ -2,6 +2,6 @@
 
 module Legion
   module LLM
-    VERSION = '0.6.0'
+    VERSION = '0.6.2'
   end
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: legion-llm
 version: !ruby/object:Gem::Version
-  version: 0.6.0
+  version: 0.6.2
 platform: ruby
 authors:
 - Esity