nitro_intelligence 2.4.0 → 2.5.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: b9041d12c07ae072e5a7a5a2c5fccbec275d260dfd2c99cba7a5bc92337eab4d
-  data.tar.gz: 496aed9a03587308b2af556fa846686edfee3e3a1aaf4a7a3a686cf4f08b322b
+  metadata.gz: 0bc1642f0475d8292a1e9611003d79e31983fb9afe2fd5e35d5ddf821a191e53
+  data.tar.gz: f61c77b0a3cd8231ba52bd7df317ca715423dc08eb72639e028febf508d52472
 SHA512:
-  metadata.gz: ffa988435a0128d981e79a94f28530d1525b82c231f44e6a3cd87f15d459bcdd27ad2ef19b57499153a0a583256bb7bec2c711f51bd3cf519d9c9827cb080937
-  data.tar.gz: 18e594810e1c4815f759375f4692cb726dbc5415c00495aa03b3943d16c77d0079aa115f5aeb9ed368d715429aaafff6967f986c4b1496984160c7cb95c8aeea
+  metadata.gz: 7603cb32f91bda923f1728c5318db1a1b0e7fb59c1a46bca3a4ac3c01567de9d850b578626dd76f05d17b3348e1cdb5495fa6e98f28a6e9c871323147f02db24
+  data.tar.gz: 251976a344210f499e8b7b0c7c05bec38461fd4d941cd852fe1d5b52bc4d69cc052f992b22fe83a0241845b8d8212773f77e2e5a0ea5addcc7c459d4d65bde2f

data/lib/nitro_intelligence/client/handlers/base_handler.rb

@@ -16,12 +16,73 @@ module NitroIntelligence
         # supplied, so cap it rather than turning a large hash into a failed request.
         MAX_SPEND_LOGS_METADATA_BYTES = 4096
 
+        # Cost the inference gateway calculated for a response.
+        # See https://docs.litellm.ai/docs/proxy/response_headers
+        #
+        # The gateway is the only component that knows which deployment actually
+        # served a request, and the same model group can be served by internal
+        # capacity or by any of several third-party providers at different rates.
+        # Recomputing cost downstream from token counts therefore means maintaining
+        # a second price table that silently drifts from the one doing the billing.
+        RESPONSE_COST_HEADER = "x-litellm-response-cost".freeze
+        RESPONSE_COST_INPUT_HEADER = "x-litellm-response-cost-input".freeze
+        RESPONSE_COST_OUTPUT_HEADER = "x-litellm-response-cost-output".freeze
+
         def initialize(client:)
           @client = client
         end
 
+        # Cost breakdown for an observation, or nil when the gateway did not price
+        # the request.
+        #
+        # A deployment the gateway has no price for sends no cost header at all
+        # rather than a zero one, so absence has to mean "unknown" here. Recording
+        # nil leaves the generation without a cost; recording 0.0 would assert the
+        # request was free and quietly understate spend for every model still
+        # awaiting a price.
+        #
+        # Only the total is guaranteed. Where the cost comes from the upstream
+        # provider rather than the gateway's own calculation - OpenRouter reports a
+        # real per-request cost of its own, which the gateway passes through - there
+        # is no component breakdown, so input and output appear only when sent.
+        def cost_details(response)
+          headers = response_headers(response)
+          return nil if headers.nil?
+
+          total = header_amount(headers, RESPONSE_COST_HEADER)
+          return nil if total.nil?
+
+          {
+            total:,
+            input: header_amount(headers, RESPONSE_COST_INPUT_HEADER),
+            output: header_amount(headers, RESPONSE_COST_OUTPUT_HEADER),
+          }.compact
+        end
+
       private
 
+        # `last_response` carries the HTTP metadata of the response a typed model was
+        # built from. The client leaves it unset on nested and locally constructed
+        # models, and on endpoints returning raw or binary payloads, so both the
+        # method and its value are optional.
+        def response_headers(response)
+          return nil unless response.respond_to?(:last_response)
+
+          response.last_response&.headers
+        end
+
+        # Header values arrive as strings. A malformed one is worth ignoring rather
+        # than raising: a trace missing its cost is a far smaller problem than an
+        # inference call failing because the gateway sent something unexpected.
+        def header_amount(headers, name)
+          value = headers[name]
+          return nil if value.blank?
+
+          Float(value)
+        rescue ArgumentError, TypeError
+          nil
+        end
+
         def add_request_headers(parameters, headers)
           request_options = (parameters[:request_options] ||= {})
           (request_options[:extra_headers] ||= {}).merge!(headers.compact)

data/lib/nitro_intelligence/client/handlers/observed/audio_transcription_handler.rb

@@ -79,6 +79,7 @@ module NitroIntelligence
                 output_tokens: audio_transcription.usage.output_tokens,
                 total_tokens: audio_transcription.usage.total_tokens,
               },
+              cost_details: @base_handler.cost_details(audio_transcription),
             }
 
             [audio_transcription, trace_attributes]

data/lib/nitro_intelligence/client/handlers/observed/chat_handler.rb

@@ -59,6 +59,7 @@ module NitroIntelligence
                 completion_tokens: chat_completion.usage.completion_tokens,
                 total_tokens: chat_completion.usage.total_tokens,
               },
+              cost_details: @base_handler.cost_details(chat_completion),
             }
 
             [chat_completion, trace_attributes]

data/lib/nitro_intelligence/client/handlers/observed/image_handler.rb

@@ -94,6 +94,7 @@ module NitroIntelligence
                 completion_tokens: chat_completion.usage.completion_tokens,
                 total_tokens: chat_completion.usage.total_tokens,
               },
+              cost_details: @base_handler.cost_details(chat_completion),
             }
 
             [chat_completion, trace_attributes]

data/lib/nitro_intelligence/client/observers/langfuse_observer.rb

@@ -42,17 +42,7 @@ module NitroIntelligence
               record_input(generation, input)
 
               result, trace_attributes = observe_failures(generation) { yield(generation) }
-
-              if trace_attributes
-                handle_truncation(trace_attributes[:input], trace_attributes[:output], trace_attributes[:model])
-
-                generation.model = trace_attributes[:model] if trace_attributes[:model]
-                generation.usage_details = trace_attributes[:usage_details] if trace_attributes[:usage_details]
-                generation.input = trace_attributes[:input] if trace_attributes[:input]
-                generation.output = trace_attributes[:output] if trace_attributes[:output]
-
-                generation.update_trace(input: trace_attributes[:input], output: trace_attributes[:output])
-              end
+              record_result(generation, trace_attributes)
 
               result
             end
@@ -61,6 +51,27 @@ module NitroIntelligence
 
       private
 
+        # Applied once the response is in hand, so the observation reflects what came
+        # back rather than what was asked for. Each attribute is set only when the
+        # handler supplied it: an observation that records nothing is better than one
+        # asserting a value the response never carried. Cost is the clearest case -
+        # the gateway omits the header entirely for a deployment it has no price for,
+        # and writing a zero there would read as a free request rather than an
+        # unpriced one.
+        def record_result(generation, trace_attributes)
+          return unless trace_attributes
+
+          handle_truncation(trace_attributes[:input], trace_attributes[:output], trace_attributes[:model])
+
+          generation.model = trace_attributes[:model] if trace_attributes[:model]
+          generation.usage_details = trace_attributes[:usage_details] if trace_attributes[:usage_details]
+          generation.cost_details = trace_attributes[:cost_details] if trace_attributes[:cost_details]
+          generation.input = trace_attributes[:input] if trace_attributes[:input]
+          generation.output = trace_attributes[:output] if trace_attributes[:output]
+
+          generation.update_trace(input: trace_attributes[:input], output: trace_attributes[:output])
+        end
+
         # Recorded before the request is made so that a request which raises still
         # shows what was sent. Handlers whose input is not safe to record twice
         # (image generation sends base64 payloads that are replaced with media

data/lib/nitro_intelligence/reporter.rb

@@ -4,6 +4,8 @@ require "uri"
 
 module NitroIntelligence
   class Reporter
+    class DatasetItemError < StandardError; end
+
     def initialize(observability_project_slug:)
       @observability_project_slug = observability_project_slug
       @project_client = fetch_project_client
@@ -20,7 +22,17 @@ module NitroIntelligence
       request["Authorization"] = "Basic #{@project_client.project.auth_token}"
       request.body = attributes.to_json
 
-      http.request(request)
+      response = http.request(request)
+
+      # Every other request in this gem raises on an unsuccessful response. Without this a
+      # rejected write - bad credentials, a malformed item, a dataset that does not exist -
+      # is indistinguishable from a successful one, and a caller building a dataset ends up
+      # with a run against items that were never stored.
+      unless response.is_a?(Net::HTTPSuccess)
+        raise DatasetItemError, "#{response.code} creating dataset item: #{response.body}"
+      end
+
+      response
     end
 
     def score(trace_id:, name:, value:, id: "#{trace_id}-#{name}")

data/lib/nitro_intelligence/version.rb

@@ -1,3 +1,3 @@
 module NitroIntelligence
-  VERSION = "2.4.0".freeze
+  VERSION = "2.5.0".freeze
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: nitro_intelligence
 version: !ruby/object:Gem::Version
-  version: 2.4.0
+  version: 2.5.0
 platform: ruby
 authors:
 - Igor Artemenko
@@ -57,14 +57,14 @@ dependencies:
     requirements:
     - - "~>"
       - !ruby/object:Gem::Version
-        version: '0.58'
+        version: '0.79'
   type: :runtime
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - "~>"
       - !ruby/object:Gem::Version
-        version: '0.58'
+        version: '0.79'
 description: The Ruby client for Nitro Intelligence
 email:
 - igor.artemenko@powerhrg.com