lex-llm-azure-foundry 0.2.13 → 0.2.15

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 76e76ee4edefeb764f3d1829a900e1bf60bf9ccc598532feb0d6033680af329f
-  data.tar.gz: 778667f3720744c2b5ea38e96d52db29d22c4228f4247a55e9976fd2e1b85612
+  metadata.gz: e53b1d826dd7b500516b5d81d4d1502f4cc0c3edbe9453fe80a2436343fdece9
+  data.tar.gz: 4b9206ecc5905a2325b4a626901ea4e2cdb5166dba0c9b679a91668a584d5fd1
 SHA512:
-  metadata.gz: f1facb1bad185f1b7d2429213869f637598802d078ea1bf453c670de3ff5426f56e763d0102e3b15ce78f4afb6bd36788bf7269918416191b0daa483e5ae6ca4
-  data.tar.gz: aa1a7c88250da3f1d676ef92fbb0f46522d44994e54d49707daad0ec38eb0e940341ea4f9b0661af9d117771b4fe5d3e4156a3c29d21a11b74446dd9fdbd9570
+  metadata.gz: 4b8fce48bc00ff40a0d8e04f731ee2773b0f70bf0fa4e78e212065b063b9631a5f1af639f2ed9e11ba73945f8af47e36008f9b4c62782fe82669f7b6e9e91c24
+  data.tar.gz: 6d1a10eb83f5d62c833c51ef3ed4499b71f0d942938cca357747fbfc4bf935a5d46be0c3a0e0f052fbe42cdfcd6d9fa32610a63967ff291ff9a6b2cb18cb0774

data/CHANGELOG.md

@@ -1,5 +1,16 @@
 # Changelog
 
+## [0.2.15] - 2026-07-09
+
+### Fixed
+- Offerings now populate `limits[:context_window]` (and `max_output_tokens`), so the router sees real capacity for Azure lanes instead of nil/unbounded. `build_offering` previously never set `limits` — deployment config context sizes landed in `metadata` and were invisible to routing (a request could then mis-route to an Azure lane the router thought had unlimited context). `context_window` is sourced from live catalog when the endpoint reports it, else per-deployment instance config (`context_window`/`max_input_tokens`), else nil (a genuine per-instance gap — never a hardcoded guess). Azure's inference-plane endpoints (`model_inference GET /info`, `openai_v1 GET /models`) do not report per-model context length, mirroring the OpenAI/Bedrock cloud providers. `Model::Info#context_length` (models API) is populated the same way.
+
+## [0.2.14] - 2026-07-03
+
+### Fixed
+- Emit relative request paths from `path_for` (no leading slash). `Connection` builds Faraday with `api_base` as the base URL; on the `openai_v1` surface that base carries the `/openai/v1` path, and a leading-slash path was treated as absolute and dropped it — 404ing live discovery (empty offerings) and chat. Paths are now relative so the base path survives on both surfaces.
+- Resolve `models_url`/`health_url` per surface: `models` on `openai_v1`, `models/info` on `model_inference`. Previously always `info`, which 404s on the `openai_v1` surface.
+
 ## [0.2.13] - 2026-06-20
 
 ### Fixed

data/lib/legion/extensions/llm/azure_foundry/provider.rb

@@ -15,6 +15,9 @@ module Legion
           DEFAULT_API_VERSION = '2024-05-01-preview'
           MODEL_INFERENCE_SURFACE = :model_inference
           OPENAI_V1_SURFACE = :openai_v1
+          # Keys a live model catalog might use to report context length. Azure's own
+          # endpoints rarely do (see deployment_limits), so this is a best-effort read.
+          CATALOG_CONTEXT_KEYS = %i[context_window max_input_tokens context_length].freeze
 
           class << self
             def slug = 'azure_foundry'
@@ -143,7 +146,7 @@ module Legion
           def completion_url = path_for('chat/completions')
           def chat_url = completion_url
           def stream_url = completion_url
-          def models_url = path_for('info')
+          def models_url = surface == MODEL_INFERENCE_SURFACE ? path_for('info') : path_for('models')
           def embedding_url(**) = path_for('embeddings')
           def health_url = models_url
 
@@ -161,7 +164,8 @@ module Legion
               model_family: normalize_family(model_family || configured_family || infer_model_family(model_id)),
               canonical_model_alias: canonical_model_alias || configured_alias,
               usage_type: usage_type || value_for(deployment, :usage_type) || usage_type_for(model_id),
-              metadata: metadata.merge(deployment_metadata(deployment))
+              metadata: metadata.merge(deployment_metadata(deployment)),
+              limits: deployment_limits(deployment)
             )
           end
 
@@ -275,6 +279,7 @@ module Legion
               provider: :azure_foundry,
               family: offering.metadata[:model_family],
               capabilities: capabilities,
+              context_length: offering.context_window,
               modalities_input: modalities[:input],
               modalities_output: modalities[:output],
               metadata: offering.to_h
@@ -285,10 +290,14 @@ module Legion
             config.azure_foundry_api_version || DEFAULT_API_VERSION
           end
 
+          # Paths MUST be relative (no leading slash). Faraday builds the
+          # connection with api_base as the base URL — on the openai_v1 surface
+          # that base carries the /openai/v1 path, and a leading-slash path would
+          # be treated as absolute and drop it, 404ing discovery and chat.
           def path_for(path)
-            prefix = surface == MODEL_INFERENCE_SURFACE ? '/models' : ''
+            prefix = surface == MODEL_INFERENCE_SURFACE ? 'models/' : ''
             suffix = surface == MODEL_INFERENCE_SURFACE ? "?api-version=#{api_version}" : ''
-            "#{prefix}/#{path}#{suffix}"
+            "#{prefix}#{path}#{suffix}"
           end
 
           def bearer_header
@@ -358,7 +367,8 @@ module Legion
             )
           end
 
-          def build_offering(model:, model_family:, usage_type:, instance_id:, canonical_model_alias:, metadata:) # rubocop:disable Metrics/ParameterLists
+          def build_offering(model:, model_family:, usage_type:, instance_id:, canonical_model_alias:, # rubocop:disable Metrics/ParameterLists
+                             metadata:, limits: {})
             policy = resolve_capability_policy(model, usage_type)
             Legion::Extensions::Llm::Routing::ModelOffering.new(
               provider_family: :azure_foundry,
@@ -369,6 +379,7 @@ module Legion
               usage_type: usage_type.to_sym,
               capabilities: policy[:capabilities],
               capability_sources: policy[:sources],
+              limits: limits,
               metadata: metadata.merge(
                 model_family: model_family,
                 canonical_model_alias: canonical_model_alias,
@@ -380,14 +391,29 @@ module Legion
           def with_live_metadata(offering)
             response = connection.get(models_url)
             metadata = offering.metadata.merge(model_info: response.body)
-            with_health(offering, ready: true, checked: true, metadata:)
+            # Prefer a context_window the live catalog actually reports; fall back to
+            # whatever the deployment config already resolved (Azure endpoints usually
+            # omit context length, so config remains authoritative).
+            limits = offering.to_h[:limits].to_h
+            catalog_window = catalog_context_window(response.body)
+            limits = limits.merge(context_window: catalog_window) if catalog_window
+            with_health(offering, ready: true, checked: true, metadata:, limits:)
           end
 
-          def with_health(offering, ready:, checked:, error: nil, metadata: offering.metadata)
+          def catalog_context_window(body)
+            return nil unless body.is_a?(Hash)
+
+            value = CATALOG_CONTEXT_KEYS.filter_map { |key| body[key] || body[key.to_s] }.first
+            value&.to_i
+          end
+
+          def with_health(offering, ready:, checked:, error: nil, metadata: offering.metadata, limits: nil) # rubocop:disable Metrics/ParameterLists
             health = { ready: ready, checked: checked }
             health = health.merge(error: error.class.name, message: error.message) if error
 
-            Legion::Extensions::Llm::Routing::ModelOffering.new(offering.to_h.merge(health:, metadata:))
+            data = offering.to_h.merge(health:, metadata:)
+            data[:limits] = limits if limits
+            Legion::Extensions::Llm::Routing::ModelOffering.new(data)
           end
 
           def filter_offerings(offerings, model_family: nil, usage_type: nil, **)
@@ -398,6 +424,24 @@ module Legion
             end
           end
 
+          # Azure's inference-plane endpoints do NOT report per-model context length
+          # (model_inference GET /info returns only model_name/model_type/
+          # model_provider_name; the openai_v1 GET /models is an OpenAI-style
+          # id/created/owned_by list). So context_window is sourced from the
+          # per-deployment instance config (keys :context_window / :max_input_tokens);
+          # live discovery (with_live_metadata) can override it when a catalog entry
+          # actually carries it. Absent both, the window is nil — a per-instance gap,
+          # never a hardcoded guess.
+          def deployment_limits(deployment)
+            return {} unless deployment
+
+            context_window = value_for(deployment, :context_window) || value_for(deployment, :max_input_tokens)
+            {
+              context_window: context_window&.to_i,
+              max_output_tokens: value_for(deployment, :max_output_tokens)&.to_i
+            }.compact
+          end
+
           def deployment_metadata(deployment)
             return {} unless deployment
 

data/lib/legion/extensions/llm/azure_foundry/version.rb

@@ -4,7 +4,7 @@ module Legion
   module Extensions
     module Llm
       module AzureFoundry
-        VERSION = '0.2.13'
+        VERSION = '0.2.15'
       end
     end
   end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: lex-llm-azure-foundry
 version: !ruby/object:Gem::Version
-  version: 0.2.13
+  version: 0.2.15
 platform: ruby
 authors:
 - LegionIO