nitro_intelligence 4.3.0 → 4.3.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
checksums.yaml CHANGED
@@ -1,7 +1,7 @@
1
1
  ---
2
2
  SHA256:
3
- metadata.gz: 27f08824317caeb5e6b8de73db8cec8bbd9633a6dca72fc96d37977f7d4776e6
4
- data.tar.gz: 2d751c4d40c30148d47bf75ac36b0b4a7024b7740b62985f79d2b2d581efd4c0
3
+ metadata.gz: 7257a5598a1874c82985d7ad53c510630324d188fb0b8357b9cbd0f2bd2a2a01
4
+ data.tar.gz: 257cf46ad150fdf356ffb18163d632f78e4c1cf83ada368fb821bac28af12e7a
5
5
  SHA512:
6
- metadata.gz: aae6508c9d5f23520d46acf912f9dcb4f18643e7950052a24c99f91acb0f99f8ef84156eea767733f1e50314b758dd10d982d8acab10de9c85e1cba0a0a4efa3
7
- data.tar.gz: 637cb07fea164e376e4c6ca24ad1642bb98e9577c6cc6ea60429c1af5d58ca1f389a8113e71143b7eb96a292d54325758daa4a67af7bbe9f741295ffce233bb8
6
+ metadata.gz: 57af436d4123e3f5ebe01c00e035d41bcc7a6de019c0318d8621b07db6dc25e780548414aaa3ce0975938ac686503db9b2ed942e9ebc2d0c3ef6fb3c41ad49a3
7
+ data.tar.gz: fbe0b679170af183511c157dd59fad3ed5e9fd13cc38bb9dfac19fa6a3fe17b7868a71181ed7dcf8cfe2ad82f38d71e8daecf48f113e96e0971741f32fd0171b
data/docs/CHANGELOG.md CHANGED
@@ -5,6 +5,15 @@ All notable changes to this project will be documented in this file.
5
5
  The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
6
6
  and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
7
7
 
8
+ ## [4.3.1](https://github.com/powerhome/nitro-intelligence.rb/compare/v4.3.0...v4.3.1) (2026-11-11)
9
+
10
+
11
+ ### Fixed
12
+
13
+ * record the cached-input and cache-write tokens a chat generation reports in Cerebro ([#211](https://github.com/powerhome/nitro-intelligence.rb/issues/211)) ([460a10f](https://github.com/powerhome/nitro-intelligence.rb/commit/460a10f2fb3b2a04ad28369eeec71ebadc3318a4))
14
+ * record the reasoning tokens a chat generation spent ([#211](https://github.com/powerhome/nitro-intelligence.rb/issues/211)) ([460a10f](https://github.com/powerhome/nitro-intelligence.rb/commit/460a10f2fb3b2a04ad28369eeec71ebadc3318a4))
15
+ * show a chat generation's reasoning in Cerebro's formatted view ([#212](https://github.com/powerhome/nitro-intelligence.rb/issues/212)) ([d7ffc8d](https://github.com/powerhome/nitro-intelligence.rb/commit/d7ffc8dc42d666af09d4f2ccade997b0345959d1))
16
+
8
17
  ## [4.3.0](https://github.com/powerhome/nitro-intelligence.rb/compare/v4.2.0...v4.3.0) (2026-11-11)
9
18
 
10
19
 
@@ -65,8 +65,95 @@ module NitroIntelligence
65
65
  }.compact
66
66
  end
67
67
 
68
+ # A chat completion's message as Cerebro should record it, with its reasoning moved to
69
+ # where Cerebro's formatted view shows it.
70
+ #
71
+ # The gateway returns reasoning as the message's `reasoning_content`, and LiteLLM repeats
72
+ # it as `provider_specific_fields.reasoning`. Cerebro's formatted view reads neither: it
73
+ # shows a message's reasoning only from its `thinking` list, so the text is otherwise
74
+ # visible in the JSON view alone, and recorded twice. Here it is recorded once, as
75
+ # `thinking`. Only reasoning that is text is moved: anything else a provider sends there,
76
+ # and every other provider field, is kept as the gateway returned it.
77
+ #
78
+ # A copy, because openai-ruby's `to_h` is the message's own hash: changing it, or letting
79
+ # the observer truncate it, would change the response the caller gets back.
80
+ def recorded_message(message)
81
+ recorded = message.to_h.deep_dup
82
+ reasoning = reasoning_text(recorded[:reasoning_content])
83
+ recorded.delete(:reasoning_content) if reasoning
84
+ reasoning = take_provider_reasoning(recorded, reasoning) || reasoning
85
+
86
+ recorded[:thinking] = [{ type: "thinking", content: reasoning }] if reasoning
87
+ recorded
88
+ end
89
+
90
+ # The breakdowns of a chat completion's usage recorded in Cerebro, each under the name
91
+ # Cerebro is sent and the names the gateway may report it under, first match winning.
92
+ # Cerebro takes every breakdown it is sent out of the input or output count and records
93
+ # it alongside, so only breakdowns that are part of that count and do not overlap one
94
+ # another are sent. The gateway reports others that are not: it mirrors a cache write
95
+ # as both `cache_write_tokens` and `cache_creation_tokens`, splits tokens by modality
96
+ # (`text_tokens`, `image_tokens` and the like) across the cached and reasoning ones,
97
+ # and adds counts such as `initial_prompt_tokens` that are not part of the input at all.
98
+ PROMPT_TOKEN_DETAILS = {
99
+ cached_tokens: %i[cached_tokens],
100
+ cache_write_tokens: %i[cache_write_tokens cache_creation_tokens],
101
+ }.freeze
102
+ COMPLETION_TOKEN_DETAILS = {
103
+ reasoning_tokens: %i[reasoning_tokens],
104
+ }.freeze
105
+
106
+ # A chat completion's token usage, in the shape the gateway reports it, breakdowns
107
+ # included. Cerebro reads this shape itself: it records each breakdown under a key of
108
+ # its own (reasoning as `output_reasoning_tokens`, cached input as
109
+ # `input_cached_tokens`) and takes it out of the output or input count it belongs to,
110
+ # so the total stays the one the gateway reported. Flattening a breakdown into a key of
111
+ # our own would instead be counted on top of the output it is part of.
112
+ def chat_usage_details(usage)
113
+ {
114
+ prompt_tokens: usage.prompt_tokens,
115
+ completion_tokens: usage.completion_tokens,
116
+ total_tokens: usage.total_tokens,
117
+ prompt_tokens_details: token_counts(usage.prompt_tokens_details, PROMPT_TOKEN_DETAILS),
118
+ completion_tokens_details: token_counts(usage.completion_tokens_details, COMPLETION_TOKEN_DETAILS),
119
+ }.compact
120
+ end
121
+
68
122
  private
69
123
 
124
+ def reasoning_text(value)
125
+ value if value.is_a?(String) && value.present?
126
+ end
127
+
128
+ # Removes LiteLLM's copy of the reasoning from the provider fields, and returns it, when it
129
+ # is text that repeats the message's own or the message has none. Reasoning that says
130
+ # something else, or is not text, is left where it is.
131
+ def take_provider_reasoning(recorded, reasoning)
132
+ provider_fields = recorded[:provider_specific_fields]
133
+ return nil unless provider_fields.is_a?(Hash)
134
+
135
+ provider_reasoning = reasoning_text(provider_fields[:reasoning])
136
+ return nil unless provider_reasoning && (reasoning.nil? || provider_reasoning == reasoning)
137
+
138
+ remaining = provider_fields.except(:reasoning)
139
+ if remaining.empty?
140
+ recorded.delete(:provider_specific_fields)
141
+ else
142
+ recorded[:provider_specific_fields] = remaining
143
+ end
144
+ provider_fields[:reasoning]
145
+ end
146
+
147
+ # The breakdowns `names` lists that `details` reports a count for. Cerebro rejects the
148
+ # whole usage if any breakdown value is not a count, so a value that is not one is left
149
+ # out rather than costing the generation its token counts.
150
+ def token_counts(details, names)
151
+ reported = details.to_h
152
+ names.to_h do |name, aliases|
153
+ [name, aliases.map { |key| reported[key] }.find { |value| value.is_a?(Integer) && value >= 0 }]
154
+ end.compact.presence
155
+ end
156
+
70
157
  # `last_response` carries the HTTP metadata of the response a model was built
71
158
  # from. The client leaves it unset on nested and locally constructed models,
72
159
  # so both the method and its value are optional.
@@ -98,16 +98,12 @@ module NitroIntelligence
98
98
  chat_completion = @base_handler.perform_request(
99
99
  parameters:, correlation_trace_id: generation.trace_id, mask_content: project.mask_content?
100
100
  )
101
- output = chat_completion.choices.first.message.to_h
101
+ output = @base_handler.recorded_message(chat_completion.choices.first.message)
102
102
 
103
103
  trace_attributes = {
104
104
  model: chat_completion.model,
105
105
  output:,
106
- usage_details: {
107
- prompt_tokens: chat_completion.usage.prompt_tokens,
108
- completion_tokens: chat_completion.usage.completion_tokens,
109
- total_tokens: chat_completion.usage.total_tokens,
110
- },
106
+ usage_details: @base_handler.chat_usage_details(chat_completion.usage),
111
107
  cost_details: @base_handler.cost_details(chat_completion),
112
108
  }
113
109
 
@@ -89,18 +89,14 @@ module NitroIntelligence
89
89
  # A copy, because recording replaces its image URLs and the caller may send the same
90
90
  # messages again.
91
91
  input = parameters[:messages].deep_dup
92
- output = chat_completion.choices.first.message.to_h
92
+ output = @base_handler.recorded_message(chat_completion.choices.first.message)
93
93
  handle_image_generation_uploads(input, output, image_generation)
94
94
 
95
95
  trace_attributes = {
96
96
  model: chat_completion.model,
97
97
  input:,
98
98
  output:,
99
- usage_details: {
100
- prompt_tokens: chat_completion.usage.prompt_tokens,
101
- completion_tokens: chat_completion.usage.completion_tokens,
102
- total_tokens: chat_completion.usage.total_tokens,
103
- },
99
+ usage_details: @base_handler.chat_usage_details(chat_completion.usage),
104
100
  cost_details: @base_handler.cost_details(chat_completion),
105
101
  }
106
102
 
@@ -1,4 +1,4 @@
1
1
  module NitroIntelligence
2
2
  # The gem's version.
3
- VERSION = "4.3.0".freeze
3
+ VERSION = "4.3.1".freeze
4
4
  end
metadata CHANGED
@@ -1,7 +1,7 @@
1
1
  --- !ruby/object:Gem::Specification
2
2
  name: nitro_intelligence
3
3
  version: !ruby/object:Gem::Version
4
- version: 4.3.0
4
+ version: 4.3.1
5
5
  platform: ruby
6
6
  authors:
7
7
  - Igor Artemenko