Module: Legion::LLM::API::Translators::OpenAIResponse

Extended by:
Legion::Logging::Helper
Defined in:
lib/legion/llm/api/translators/openai_response.rb

Constant Summary collapse

FINISH_REASON_MAP = 
{
  'stop'           => 'stop',
  'length'         => 'length',
  'tool_calls'     => 'tool_calls',
  'content_filter' => 'content_filter'
}.freeze

Class Method Summary collapse

Class Method Details

.build_tool_calls(pipeline_response) ⇒ Object 



178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
 
# File 'lib/legion/llm/api/translators/openai_response.rb', line 178

def build_tool_calls(pipeline_response)
  tools_data = pipeline_response.respond_to?(:tools) ? pipeline_response.tools : nil
  return [] unless tools_data.is_a?(Array) && !tools_data.empty?

  tools_data.each_with_index.filter_map do |tc, idx|
    name = tc.respond_to?(:name) ? tc.name : (tc[:name] || tc['name'])
    args = tc.respond_to?(:arguments) ? tc.arguments : (tc[:arguments] || tc['arguments'] || {})
    tc_id = tc.respond_to?(:id) ? tc.id : (tc[:id] || tc['id'] || "call_#{SecureRandom.hex(8)}")
    next unless name

    {
      id:       tc_id,
      type:     'function',
      index:    idx,
      function: {
        name:      name.to_s,
        arguments: args.is_a?(String) ? args : Legion::JSON.dump(args)
      }
    }
  end
end

.content_looks_like_tool_json?(content) ⇒ Boolean

Heuristic: does the content look like a bare JSON object that is tool-call arguments (e.g. “…”, “limit”: 300)?

Returns:

  • (Boolean) 


228
229
230
231
232
233
234
235
236
 
# File 'lib/legion/llm/api/translators/openai_response.rb', line 228

def content_looks_like_tool_json?(content)
  stripped = content.to_s.strip
  return false unless stripped.start_with?('{"') && stripped.end_with?('}')

  parsed = Legion::JSON.parse(stripped, symbolize_names: false)
  parsed.is_a?(Hash) && parsed.keys.any?
rescue Legion::JSON::ParseError, StandardError
  false
end

.embedding_token_count(usage, input_text) ⇒ Object 



216
217
218
219
220
221
222
223
224
 
# File 'lib/legion/llm/api/translators/openai_response.rb', line 216

def embedding_token_count(usage, input_text)
  usage_hash = usage.respond_to?(:key?) ? usage : {}
  token_count = usage_hash[:prompt_tokens] || usage_hash['prompt_tokens'] ||
                usage_hash[:input_tokens] || usage_hash['input_tokens'] ||
                usage_hash[:total_tokens] || usage_hash['total_tokens']
  return token_count.to_i if token_count

  input_text.to_s.split.size
end

.extract_token_count(tokens, key) ⇒ Object 



206
207
208
209
210
211
212
213
214
 
# File 'lib/legion/llm/api/translators/openai_response.rb', line 206

def extract_token_count(tokens, key)
  return nil if tokens.nil?
  return tokens[key] || tokens[key.to_s] if tokens.is_a?(Hash)

  method_name = { input: :input_tokens, output: :output_tokens }[key]
  return tokens.public_send(method_name) if method_name && tokens.respond_to?(method_name)

  nil
end

.format_chat_completion(pipeline_response, model:, request_id: nil, include_reasoning: false) ⇒ Object 



23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
 
# File 'lib/legion/llm/api/translators/openai_response.rb', line 23

def format_chat_completion(pipeline_response, model:, request_id: nil, include_reasoning: false)
  request_id ||= SecureRandom.uuid
  routing = pipeline_response.routing || {}
  tokens = pipeline_response.tokens || {}
  raw_msg = pipeline_response.message
  content = raw_msg.is_a?(Hash) ? (raw_msg[:content] || raw_msg['content']) : raw_msg.to_s
  stop_reason = pipeline_response.stop&.dig(:reason)&.to_s
  tool_calls = build_tool_calls(pipeline_response)
  resolved_model = (routing[:model] || routing['model'] || model).to_s

  log.debug("[llm][translator][openai_response] action=format_chat_completion request_id=#{request_id} model=#{resolved_model}")

  finish_reason = tool_calls.empty? ? map_finish_reason(stop_reason) : 'tool_calls'

  # When tool calls are present and content is just JSON arguments
  # (e.g. vLLM/qwen forced tool choice), clear the content field
  # so the client sees only structured tool_calls.
  content = nil if tool_calls.any? && content_looks_like_tool_json?(content)

  message_body = { role: 'assistant', content: content }
  message_body[:tool_calls] = tool_calls unless tool_calls.empty?

  # Include reasoning/thinking content in the response when requested.
  # Uses the `reasoning_content` field convention from OpenAI's reasoning models.
  if include_reasoning && pipeline_response.respond_to?(:thinking) && pipeline_response.thinking
    thinking_data = pipeline_response.thinking
    reasoning_text = if thinking_data.is_a?(Hash)
                       thinking_data[:content] || thinking_data['content'] || thinking_data[:text] || thinking_data['text']
                     elsif thinking_data.respond_to?(:content)
                       thinking_data.content
                     elsif thinking_data.respond_to?(:text)
                       thinking_data.text
                     else
                       thinking_data.to_s
                     end
    message_body[:reasoning_content] = reasoning_text.to_s unless reasoning_text.to_s.empty?
  end

  {
    id:      "chatcmpl-#{request_id.delete('-')}",
    object:  'chat.completion',
    created: Time.now.to_i,
    model:   resolved_model,
    choices: [
      {
        index:         0,
        message:       message_body,
        finish_reason: finish_reason
      }
    ],
    usage:   {
      prompt_tokens:     extract_token_count(tokens, :input),
      completion_tokens: extract_token_count(tokens, :output),
      total_tokens:      (extract_token_count(tokens, :input).to_i + extract_token_count(tokens, :output).to_i)
    }
  }
end

.format_embeddings(vector, model:, input_text:, usage: nil) ⇒ Object 



141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
 
# File 'lib/legion/llm/api/translators/openai_response.rb', line 141

def format_embeddings(vector, model:, input_text:, usage: nil)
  tokens = embedding_token_count(usage, input_text)

  {
    object: 'list',
    data:   [
      {
        object:    'embedding',
        embedding: vector,
        index:     0
      }
    ],
    model:  model.to_s,
    usage:  {
      prompt_tokens: tokens,
      total_tokens:  tokens
    }
  }
end

.format_model_object(id, created: nil, owned_by: 'legion', limits: nil) ⇒ Object 



161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
 
# File 'lib/legion/llm/api/translators/openai_response.rb', line 161

def format_model_object(id, created: nil, owned_by: 'legion', limits: nil)
  obj = {
    id:       id.to_s,
    object:   'model',
    created:  created || Time.now.to_i,
    owned_by: owned_by
  }
  if limits.is_a?(Hash)
    if limits[:context_window]
      obj[:context_window] = limits[:context_window]
      obj[:context_size] = limits[:context_window]
    end
    obj[:max_output_tokens] = limits[:max_output_tokens] if limits[:max_output_tokens]
  end
  obj
end

.format_stream_chunk(delta_text, model:, request_id:, finish_reason: nil, usage: nil) ⇒ Object 



81
82
83
84
85
86
87
88
89
90
91
92
93
94
 
# File 'lib/legion/llm/api/translators/openai_response.rb', line 81

def format_stream_chunk(delta_text, model:, request_id:, finish_reason: nil, usage: nil)
  choice = { index: 0, delta: {}, finish_reason: finish_reason }
  choice[:delta][:content] = delta_text if delta_text && !delta_text.empty?

  chunk = {
    id:      "chatcmpl-#{request_id.delete('-')}",
    object:  'chat.completion.chunk',
    created: Time.now.to_i,
    model:   model.to_s,
    choices: [choice]
  }
  chunk[:usage] = usage if usage
  chunk
end

.format_stream_delta_chunk(delta, model:, request_id:, finish_reason: nil) ⇒ Object 



125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
 
# File 'lib/legion/llm/api/translators/openai_response.rb', line 125

def format_stream_delta_chunk(delta, model:, request_id:, finish_reason: nil)
  {
    id:      "chatcmpl-#{request_id.delete('-')}",
    object:  'chat.completion.chunk',
    created: Time.now.to_i,
    model:   model.to_s,
    choices: [
      {
        index:         0,
        delta:         delta,
        finish_reason: finish_reason
      }
    ]
  }
end

.format_stream_tool_call_chunk(tool_call, model:, request_id:, index:) ⇒ Object 



96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
 
# File 'lib/legion/llm/api/translators/openai_response.rb', line 96

def format_stream_tool_call_chunk(tool_call, model:, request_id:, index:)
  fn = tool_call.is_a?(Hash) ? (tool_call[:function] || tool_call['function'] || {}) : {}
  name = tool_call.respond_to?(:name) ? tool_call.name : (tool_call[:name] || tool_call['name'] || fn[:name] || fn['name'])
  args = if tool_call.respond_to?(:arguments)
           tool_call.arguments
         else
           tool_call[:arguments] || tool_call['arguments'] || fn[:arguments] || fn['arguments'] || {}
         end
  tc_id = tool_call.respond_to?(:id) ? tool_call.id : (tool_call[:id] || tool_call['id'] || "call_#{SecureRandom.hex(8)}")

  format_stream_delta_chunk(
    {
      tool_calls: [
        {
          index:    index,
          id:       tc_id,
          type:     'function',
          function: {
            name:      name.to_s,
            arguments: args.is_a?(String) ? args : Legion::JSON.dump(args)
          }
        }
      ]
    },
    model:      model,
    request_id: request_id
  )
end

.map_finish_reason(stop_reason) ⇒ Object 



200
201
202
203
204
 
# File 'lib/legion/llm/api/translators/openai_response.rb', line 200

def map_finish_reason(stop_reason)
  return 'stop' if stop_reason.nil? || stop_reason.to_s.empty?

  FINISH_REASON_MAP.fetch(stop_reason.to_s, 'error')
end