Module: Tracekit::LLM::OpenAIInstrumentation

Defined in:
lib/tracekit/llm/openai_instrumentation.rb

Defined Under Namespace

Classes: OpenAIStreamAccumulator

Class Method Summary collapse

Class Method Details

.install(tracer) ⇒ Object 



10
11
12
13
14
15
16
17
18
19
20
21
22
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
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
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
124
125
126
127
128
129
 
# File 'lib/tracekit/llm/openai_instrumentation.rb', line 10

def install(tracer)
  # Try to load the OpenAI gem
  begin
    require "openai"
  rescue LoadError
    # openai gem not available, check if it's already defined (e.g. in tests)
    return false unless defined?(::OpenAI::Client)
  end

  client_class = ::OpenAI::Client
  return false unless client_class

  # Create the prepend module dynamically with tracer closure
  instrumentation_mod = Module.new do
    define_method(:chat) do |parameters: {}|
      model = parameters[:model] || parameters["model"] || "unknown"
      stream_proc = parameters[:stream] || parameters["stream"]
      is_streaming = stream_proc.is_a?(Proc)
      capture = Common.capture_content?

      span = tracer.start_span("chat #{model}", kind: :client)

      begin
        Common.set_request_attributes(span,
          provider: "openai",
          model: model,
          max_tokens: parameters[:max_tokens] || parameters["max_tokens"] || parameters[:max_completion_tokens] || parameters["max_completion_tokens"],
          temperature: parameters[:temperature] || parameters["temperature"],
          top_p: parameters[:top_p] || parameters["top_p"]
        )

        # Capture input content
        if capture
          messages = parameters[:messages] || parameters["messages"]
          if messages
            system_msgs = messages.select { |m| (m[:role] || m["role"]) == "system" }
            non_system = messages.reject { |m| (m[:role] || m["role"]) == "system" }
            Common.capture_system_instructions(span, system_msgs) if system_msgs.any?
            Common.capture_input_messages(span, non_system)
          end
        end

        if is_streaming
          # ruby-openai handles streaming via proc callback internally.
          # The chat method returns the final response hash, not an enumerator.
          # We wrap the user's proc to accumulate span data from each chunk.
          accumulator = OpenAIStreamAccumulator.new(span, capture)
          wrapper_proc = proc do |chunk, bytesize|
            accumulator.process_chunk(chunk)
            # Call original proc with same args
            if stream_proc.arity == 2 || stream_proc.arity < 0
              stream_proc.call(chunk, bytesize)
            else
              stream_proc.call(chunk)
            end
          end

          # Inject stream_options.include_usage for token counting
          params = parameters.dup
          so = params[:stream_options] || params["stream_options"] || {}
          unless so[:include_usage] || so["include_usage"]
            params[:stream_options] = so.merge(include_usage: true)
          end
          params[:stream] = wrapper_proc

          result = super(parameters: params)
          accumulator.finalize
          result
        else
          result = super(parameters: parameters)

          # Non-streaming response handling
          handle_response(span, result, capture)
          result
        end
      rescue => e
        Common.set_error_attributes(span, e)
        span.finish
        raise
      end
    end

    private

    def handle_response(span, result, capture)
      choices = result.dig("choices") || []
      Common.set_response_attributes(span,
        model: result["model"],
        id: result["id"],
        finish_reasons: choices.map { |c| c["finish_reason"] }.compact,
        input_tokens: result.dig("usage", "prompt_tokens"),
        output_tokens: result.dig("usage", "completion_tokens")
      )

      # Tool calls
      choices.each do |choice|
        (choice.dig("message", "tool_calls") || []).each do |tc|
          Common.record_tool_call(span,
            name: tc.dig("function", "name") || "unknown",
            id: tc["id"],
            arguments: tc.dig("function", "arguments")
          )
        end
      end

      # Output content capture
      if capture && choices.any?
        output_msgs = choices.map { |c| c["message"] }.compact
        Common.capture_output_messages(span, output_msgs) if output_msgs.any?
      end
    rescue => _e
      # Never break user code
    ensure
      span.finish
    end
  end

  client_class.prepend(instrumentation_mod)
  true
end