Module: Legion::LLM::Inference

Extended by:
Legion::Logging::Helper
Defined in:
lib/legion/llm/inference.rb,
lib/legion/llm/inference/steps.rb,
lib/legion/llm/inference/prompt.rb,
lib/legion/llm/inference/profile.rb,
lib/legion/llm/inference/request.rb,
lib/legion/llm/inference/tracing.rb,
lib/legion/llm/inference/executor.rb,
lib/legion/llm/inference/response.rb,
lib/legion/llm/inference/timeline.rb,
lib/legion/llm/inference/steps/rbac.rb,
lib/legion/llm/inference/gaia_caller.rb,
lib/legion/llm/inference/conversation.rb,
lib/legion/llm/inference/steps/debate.rb,
lib/legion/llm/inference/steps/billing.rb,
lib/legion/llm/inference/steps/logging.rb,
lib/legion/llm/inference/route_attempts.rb,
lib/legion/llm/inference/steps/metering.rb,
lib/legion/llm/inference/audit_publisher.rb,
lib/legion/llm/inference/steps/rag_guard.rb,
lib/legion/llm/inference/tool_dispatcher.rb,
lib/legion/llm/inference/native_tool_loop.rb,
lib/legion/llm/inference/steps/tool_calls.rb,
lib/legion/llm/inference/steps/rag_context.rb,
lib/legion/llm/inference/steps/prompt_cache.rb,
lib/legion/llm/inference/steps/token_budget.rb,
lib/legion/llm/inference/steps/tool_history.rb,
lib/legion/llm/inference/enrichment_injector.rb,
lib/legion/llm/inference/steps/gaia_advisory.rb,
lib/legion/llm/inference/steps/mcp_discovery.rb,
lib/legion/llm/inference/steps/post_response.rb,
lib/legion/llm/inference/steps/tier_assigner.rb,
lib/legion/llm/inference/steps/trigger_match.rb,
lib/legion/llm/inference/steps/classification.rb,
lib/legion/llm/inference/steps/skill_injector.rb,
lib/legion/llm/inference/steps/span_annotator.rb,
lib/legion/llm/inference/steps/sticky_helpers.rb,
lib/legion/llm/inference/steps/sticky_persist.rb,
lib/legion/llm/inference/steps/sticky_runners.rb,
lib/legion/llm/inference/steps/tool_discovery.rb,
lib/legion/llm/inference/steps/knowledge_capture.rb,
lib/legion/llm/inference/steps/confidence_scoring.rb

Defined Under Namespace

Modules: AuditPublisher, Conversation, EnrichmentInjector, GaiaCaller, NativeToolLoop, Profile, Prompt, RouteAttempts, Steps, Tracing Classes: Executor, Request, Response, Timeline

Constant Summary collapse

FRAMEWORK_KEYS = 
%i[request_id source timestamp datetime task_id parent_id master_id
check_subtask generate_task catch_exceptions worker_id principal_id
principal_type caller].freeze
AUTO_ROUTING_MODEL_KEY = 
'legionio'
AUTO_ROUTING_MODEL_ALIASES = 
[AUTO_ROUTING_MODEL_KEY].freeze
ToolDispatcher = 
Tools::Dispatcher

Class Method Summary collapse

Class Method Details

.apply_response_guards(result, kwargs) ⇒ Object 



737
738
739
740
741
742
743
744
745
746
747
748
749
750
 
# File 'lib/legion/llm/inference.rb', line 737

def apply_response_guards(result, kwargs)
  context = kwargs[:context]
  response_text = result[:response] || result[:content]
  guard_result = Hooks::ResponseGuard.guard_response(
    response: response_text, context: context
  )

  log.warn "[llm][inference] response_guard passed=#{guard_result[:passed]}" unless guard_result[:passed]

  result.merge(_guard_result: guard_result)
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.inference.apply_response_guards')
  result
end

.ask(message:, model: nil, provider: nil, intent: nil, tier: nil, context: {}, identity: nil) ⇒ Object 



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
 
# File 'lib/legion/llm/inference.rb', line 78

def ask(message:, model: nil, provider: nil, intent: nil, tier: nil,
        context: {}, identity: nil, &)
  started_at = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
  log_inference_request(
    request_type:       :ask,
    requested_model:    model,
    requested_provider: provider,
    intent:             intent,
    tier:               tier,
    message:            message,
    kwargs:             { context: context, identity: identity }
  )

  if Call::DaemonClient.available?
    result = daemon_ask(message: message, model: model, provider: provider,
                        context: context, tier: tier, identity: identity)
    if result
      log_inference_response(
        request_type:       :ask,
        requested_model:    model,
        requested_provider: provider,
        result:             result,
        duration_ms:        elapsed_ms_since(started_at)
      )
      return result
    end
  end

  result = ask_direct(message: message, model: model, provider: provider,
                      intent: intent, tier: tier, &)
  log_inference_response(
    request_type:       :ask,
    requested_model:    model,
    requested_provider: provider,
    result:             result,
    duration_ms:        elapsed_ms_since(started_at)
  )
  result
rescue StandardError => e
  log_inference_error(
    request_type:       :ask,
    requested_model:    model,
    requested_provider: provider,
    error:              e,
    duration_ms:        elapsed_ms_since(started_at)
  )
  raise
end

.ask_direct(message:, model: nil, provider: nil, intent: nil, tier: nil) ⇒ Object 



421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
 
# File 'lib/legion/llm/inference.rb', line 421

def ask_direct(message:, model: nil, provider: nil, intent: nil, tier: nil, &)
  assert_external_allowed! if effective_tier_is_external?(tier, provider)
  result = chat_direct(
    model:    model,
    provider: provider,
    intent:   intent,
    tier:     tier,
    message:  message,
    &
  )
  return result if result.is_a?(Hash) && result[:deferred]
  return normalize_ask_direct_hash(result, fallback_model: model || Legion::Settings[:llm][:default_model]) if result.is_a?(Hash)

  response, resolved_model = resolve_ask_direct_response(result, message, model, &)

  {
    status:   :done,
    response: response.content,
    meta:     {
      tier:       :direct,
      model:      resolved_model,
      tokens_in:  response.respond_to?(:input_tokens) ? response.input_tokens : nil,
      tokens_out: response.respond_to?(:output_tokens) ? response.output_tokens : nil
    }
  }
end

.assert_external_allowed! ⇒ Object Also known as: assert_cloud_allowed!

Raises:



844
845
846
847
848
849
850
851
 
# File 'lib/legion/llm/inference.rb', line 844

def assert_external_allowed!
  return unless enterprise_privacy?

  emit_privacy_blocked_audit
  raise Legion::LLM::PrivacyModeError,
        'External LLM tiers are disabled: enterprise_data_privacy is enabled. ' \
        'Only local and fleet tiers are permitted.'
end

.attach_escalation_history(response, history, resolution, chain) ⇒ Object 



657
658
659
660
661
662
663
664
 
# File 'lib/legion/llm/inference.rb', line 657

def attach_escalation_history(response, history, resolution, chain)
  return unless response.respond_to?(:extend)

  response.extend(EscalationHistory)
  history.each { |h| response.record_escalation_attempt(**h) }
  response.final_resolution = resolution
  response.escalation_chain = chain
end

.blocked_hook_response(blocked) ⇒ Object 



732
733
734
735
 
# File 'lib/legion/llm/inference.rb', line 732

def blocked_hook_response(blocked)
  reason = blocked[:reason] || blocked['reason'] || blocked[:message] || blocked['message'] || 'request blocked by hook'
  { error: 'request_blocked', message: reason.to_s }
end

.build_attempt(resolution, outcome, failures, duration_ms) ⇒ Object 



649
650
651
652
653
654
655
 
# File 'lib/legion/llm/inference.rb', line 649

def build_attempt(resolution, outcome, failures, duration_ms)
  attempt = { model: resolution.model, provider: resolution.provider, tier: resolution.tier,
              outcome: outcome, failures: failures, duration_ms: duration_ms }
  attempt[:offering_id] = resolution.offering_id if resolution.offering_id
  attempt[:offering_metadata] = resolution.offering_metadata unless resolution.offering_metadata.empty?
  attempt
end

.build_cache_key(model, provider, message, temperature) ⇒ Object 



757
758
759
760
761
762
763
764
765
 
# File 'lib/legion/llm/inference.rb', line 757

def build_cache_key(model, provider, message, temperature)
  messages_arr = message.is_a?(Array) ? message : [{ role: 'user', content: message.to_s }]
  Cache.key(
    model:       model || Legion::Settings[:llm][:default_model],
    provider:    provider || Legion::Settings[:llm][:default_provider],
    messages:    messages_arr,
    temperature: temperature
  )
end

.cacheable?(cache_opt, temperature, message) ⇒ Boolean

Returns:

  • (Boolean) 


752
753
754
755
 
# File 'lib/legion/llm/inference.rb', line 752

def cacheable?(cache_opt, temperature, message)
  effective_temp = temperature.nil? ? Legion::Settings[:llm][:default_temperature] : temperature
  cache_opt != false && effective_temp.to_f.zero? && message && Cache.enabled?
end

.caller_descriptor(caller_context) ⇒ Object 



331
332
333
334
335
336
337
338
339
340
 
# File 'lib/legion/llm/inference.rb', line 331

def caller_descriptor(caller_context)
  return caller_context unless caller_context.is_a?(Hash)

  source = caller_context[:source] || caller_context['source']
  path = caller_context[:path] || caller_context['path']
  return "#{source}:#{path}" if source && path
  return source.to_s if source

  caller_context.inspect
end

.chat(model: nil, provider: nil, intent: nil, tier: nil, escalate: nil, max_escalations: nil, quality_check: nil, message: nil, **kwargs) ⇒ Object

Public inference entry points — these are the methods delegated from Legion::LLM



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
 
# File 'lib/legion/llm/inference.rb', line 31

def chat(model: nil, provider: nil, intent: nil, tier: nil, escalate: nil,
         max_escalations: nil, quality_check: nil, message: nil, **kwargs, &)
  started_at = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
  log_inference_request(
    request_type:       :chat,
    requested_model:    model,
    requested_provider: provider,
    intent:             intent,
    tier:               tier,
    message:            message,
    kwargs:             kwargs
  )

  result = if defined?(Legion::Telemetry::OpenInference)
             Legion::Telemetry::OpenInference.llm_span(
               model:    (model || Legion::Settings[:llm][:default_model]).to_s,
               provider: provider&.to_s,
               input:    message
             ) do |_span|
               dispatch_chat(model: model, provider: provider, intent: intent, tier: tier, escalate: escalate,
                             max_escalations: max_escalations, quality_check: quality_check, message: message, **kwargs, &)
             end
           else
             dispatch_chat(model: model, provider: provider, intent: intent, tier: tier,
                           escalate: escalate, max_escalations: max_escalations,
                           quality_check: quality_check, message: message, **kwargs, &)
           end

  log_inference_response(
    request_type:       :chat,
    requested_model:    model,
    requested_provider: provider,
    result:             result,
    duration_ms:        elapsed_ms_since(started_at)
  )
  result
rescue StandardError => e
  log_inference_error(
    request_type:       :chat,
    requested_model:    model,
    requested_provider: provider,
    error:              e,
    duration_ms:        elapsed_ms_since(started_at)
  )
  raise
end

.chat_direct(model: nil, provider: nil, intent: nil, tier: nil, escalate: nil, max_escalations: nil, quality_check: nil, message: nil, **kwargs) ⇒ Object 



127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
 
# File 'lib/legion/llm/inference.rb', line 127

def chat_direct(model: nil, provider: nil, intent: nil, tier: nil, escalate: nil,
                max_escalations: nil, quality_check: nil, message: nil, **kwargs, &)
  log.debug(
    "[llm][inference] chat_direct.enter model=#{model} provider=#{provider} intent=#{intent} " \
    "tier=#{tier} escalate=#{escalate} message_present=#{!message.nil?} kwargs=#{kwargs.keys.sort}"
  )
  cache_opt = kwargs.delete(:cache) { true }
  temperature = kwargs.delete(:temperature)

  escalate = escalation_enabled? if escalate.nil?
  cache_key = build_cache_key(model, provider, message, temperature) if cacheable?(cache_opt, temperature, message)

  if cache_key
    cached = Cache.get(cache_key)
    if cached
      log.debug '[llm][inference] chat_direct cache=hit'
      cached_response = cached.dup
      cached_response[:meta] = (cached_response[:meta] || {}).merge(cached: true)
      return cached_response
    end
  end

  urgency = kwargs.delete(:urgency) { :normal }
  deferred = try_defer(intent: intent, urgency: urgency, model: model, provider: provider, message: message, **kwargs)
  return deferred if deferred

  log.debug(
    "[llm][inference] chat_direct.dispatch model=#{model} provider=#{provider} " \
    "escalate=#{escalate} message_present=#{!message.nil?}"
  )
  result = if escalate && message
             chat_with_escalation(
               model: model, provider: provider, intent: intent, tier: tier,
               max_escalations: max_escalations, quality_check: quality_check,
               message: message, temperature: temperature, **kwargs
             )
           else
             chat_single(model: model, provider: provider, intent: intent, tier: tier,
                         temperature: temperature, message: message, **kwargs, &)
           end
  log.debug("[llm][inference] chat_direct.exit result_class=#{result.class} result_nil=#{result.nil?}")

  if cache_key && result.is_a?(Hash)
    ttl = Legion::Settings.dig(:llm, :prompt_caching, :response_cache, :ttl_seconds) || Cache::DEFAULT_TTL
    Cache.set(cache_key, result, ttl: ttl)
  end

  result
end

.chat_single(model:, provider:, intent:, tier:, message: nil, **kwargs) ⇒ Object 



480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
 
# File 'lib/legion/llm/inference.rb', line 480

def chat_single(model:, provider:, intent:, tier:, message: nil, **kwargs, &)
  explicit_tools = kwargs.delete(:tools)
  tools = explicit_tools
  tools = nil if tools.respond_to?(:empty?) && tools.empty?

  if (intent || tier) && Router.routing_enabled?
    resolution = Router.resolve(intent: intent, tier: tier, model: model, provider: provider)
    if resolution
      model = resolution.model
      provider = resolution.provider
      assert_external_allowed! if resolution.external?
    end
  elsif tier
    assert_external_allowed! if external_tier?(tier.to_sym)
  end

  model ||= Legion::Settings[:llm][:default_model]
  instance = resolution&.instance || kwargs[:instance] || kwargs[:instance_id] || kwargs[:provider_instance]
  provider ||= (model && Router.infer_provider_for_model(model)) ||
               Legion::Settings[:llm][:default_provider]

  opts = {}
  opts[:model] = model if model
  opts[:provider] = provider if provider
  opts.merge!(kwargs.except(*FRAMEWORK_KEYS))
  opts.delete(:temperature) if opts[:temperature].nil?

  Call::Providers.inject_anthropic_cache_control!(opts, provider)
  opts[:tools] = tools if tools

  log.debug "[llm][inference] chat_single model=#{opts[:model]} provider=#{opts[:provider]} message_present=#{!message.nil?} tools=#{tools&.size || 0}"
  chat_single_native(model: opts[:model], provider: opts[:provider], instance: instance, message: message,
                     caller: kwargs[:caller], **opts.except(:model, :provider), &)
end

.chat_single_native(model:, provider:, message:, instance: nil, caller: nil, &block) ⇒ Object 



515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
 
# File 'lib/legion/llm/inference.rb', line 515

def chat_single_native(model:, provider:, message:, instance: nil, caller: nil, **, &block)
  raise native_provider_error('session-style chat requires message or messages') unless message
  raise native_provider_error('chat without a native provider') unless provider

  messages = message.is_a?(Array) ? message : [{ role: 'user', content: message.to_s }]
  result = if block
             Call::Dispatch.call(provider: provider, instance: instance, capability: :stream, model: model,
                                 messages: messages, **, &block)
           else
             Call::Dispatch.call(provider: provider, instance: instance, capability: :chat, model: model,
                                 messages: messages, **)
           end
  response = Call::NativeResponseAdapter.new(result)
  emit_non_pipeline_metering(response, model: model, provider: provider, caller: caller)
  response
end

.chat_via_pipeline(&block) ⇒ Object 



397
398
399
400
401
 
# File 'lib/legion/llm/inference.rb', line 397

def chat_via_pipeline(**, &block)
  request = Request.from_chat_args(**)
  executor = Executor.new(request)
  block ? executor.call_stream(&block) : executor.call
end

.chat_with_escalation(model:, provider:, intent:, tier:, max_escalations:, quality_check:, message:, **kwargs) ⇒ Object

Raises:



562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
 
# File 'lib/legion/llm/inference.rb', line 562

def chat_with_escalation(model:, provider:, intent:, tier:, max_escalations:, quality_check:, message:, **kwargs)
  log.debug "[llm][inference] chat_with_escalation.enter model=#{model} provider=#{provider} max_escalations=#{max_escalations}"
  chain = Router.resolve_chain(
    intent: intent, tier: tier, model: model, provider: provider,
    max_escalations: max_escalations
  )

  threshold = escalation_quality_threshold
  history = []
  last_error = nil

  chain.each do |resolution|
    response, error = run_escalation_attempt(
      resolution, message: message, kwargs: kwargs, threshold: threshold,
      quality_check: quality_check, history: history, chain: chain
    )
    last_error = error if error
    return response if response
  end

  publish_escalation_event(history, :exhausted, caller: kwargs[:caller]) if history.size > 1
  message = "All #{history.size} escalation attempts failed"
  if last_error
    providers = history.filter_map { |attempt| attempt[:provider] }.uniq.join(', ')
    message = "#{message}; no usable native provider handled the request. " \
              "providers=#{providers} last_error=#{last_error.class}: #{last_error.message}"
  end

  raise Legion::LLM::EscalationExhausted, message
end

.daemon_ask(message:, model: nil, provider: nil, context: {}, tier: nil, identity: nil) ⇒ Object

rubocop:disable Lint/UnusedMethodArgument



403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
 
# File 'lib/legion/llm/inference.rb', line 403

def daemon_ask(message:, model: nil, provider: nil, context: {}, tier: nil, identity: nil) # rubocop:disable Lint/UnusedMethodArgument
  result = Call::DaemonClient.chat(
    message: message, model: model, provider: provider,
    context: context, tier_preference: tier || :auto
  )

  case result[:status]
  when :immediate, :created
    result[:body]
  when :accepted
    Cache::Response.poll(result[:request_id])
  when :denied
    raise Legion::LLM::DaemonDeniedError, result.dig(:error, :message) || 'Access denied'
  when :rate_limited
    raise Legion::LLM::DaemonRateLimitedError, "Rate limited. Retry after #{result[:retry_after]}s"
  end
end

.direct_chat_session?(result) ⇒ Boolean

Returns:

  • (Boolean) 


448
449
450
 
# File 'lib/legion/llm/inference.rb', line 448

def direct_chat_session?(result)
  result.respond_to?(:ask) && result.respond_to?(:model) && !result.respond_to?(:content)
end

.dispatch_chat(model:, provider:, intent:, tier:, escalate:, max_escalations:, quality_check:, message:, **kwargs) ⇒ Object 



342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
 
# File 'lib/legion/llm/inference.rb', line 342

def dispatch_chat(model:, provider:, intent:, tier:, escalate:, max_escalations:, quality_check:, message:, **kwargs, &)
  log.debug(
    "[llm][inference] dispatch_chat.enter model=#{model} provider=#{provider} intent=#{intent} " \
    "tier=#{tier} escalate=#{escalate} max_escalations=#{max_escalations} " \
    "quality_check=#{quality_check} message_present=#{!message.nil?} kwargs=#{kwargs.keys.sort}"
  )
  if pipeline_enabled? && (message || kwargs[:messages]) && !block_given?
    return Prompt.dispatch(
      message || kwargs[:messages],
      intent: intent, tier: tier, provider: provider, model: model,
      escalate: escalate, max_escalations: max_escalations,
      quality_check: quality_check, **kwargs.except(:messages)
    )
  end

  if pipeline_enabled? && (message || kwargs[:messages]) && block_given?
    return chat_via_pipeline(model: model, provider: provider, intent: intent, tier: tier,
                             message: message, escalate: escalate, max_escalations: max_escalations,
                             quality_check: quality_check, **kwargs, &)
  end

  if block_given? && message
    return chat_single(model: model, provider: provider, intent: intent, tier: tier,
                       message: message, **kwargs, &)
  end

  messages = message.is_a?(Array) ? message : [{ role: 'user', content: message.to_s }]
  resolved_model = model || Legion::Settings[:llm][:default_model]

  if defined?(Legion::LLM::Hooks)
    blocked = Legion::LLM::Hooks.run_before(messages: messages, model: resolved_model)
    return blocked[:response] || blocked_hook_response(blocked) if blocked
  end

  result = chat_direct(model: model, provider: provider, intent: intent, tier: tier,
                       escalate: escalate, max_escalations: max_escalations,
                       quality_check: quality_check, message: message, **kwargs)

  if defined?(Legion::LLM::Hooks)
    blocked = Legion::LLM::Hooks.run_after(response: result, messages: messages, model: resolved_model)
    return blocked[:response] || blocked_hook_response(blocked) if blocked
  end

  result = apply_response_guards(result, kwargs) if response_guards_enabled? && result.is_a?(Hash)
  log.debug("[llm][inference] dispatch_chat.exit result_class=#{result.class} result_nil=#{result.nil?}")
  result
end

.effective_tier_is_external?(tier, provider) ⇒ Boolean Also known as: effective_tier_is_cloud?

Returns:

  • (Boolean) 


855
856
857
858
859
860
861
862
 
# File 'lib/legion/llm/inference.rb', line 855

def effective_tier_is_external?(tier, provider)
  return external_tier?(tier.to_sym) if tier
  return false unless enterprise_privacy?

  resolved = provider || Legion::Settings[:llm][:default_provider]
  external_providers = %i[anthropic bedrock openai gemini azure]
  external_providers.include?(resolved&.to_sym)
end

.elapsed_ms_since(started_at) ⇒ Object 



237
238
239
 
# File 'lib/legion/llm/inference.rb', line 237

def elapsed_ms_since(started_at)
  ((::Process.clock_gettime(::Process::CLOCK_MONOTONIC) - started_at) * 1000).round
end

.emit_non_pipeline_metering(response, model:, provider:, caller: nil) ⇒ Object 



783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
 
# File 'lib/legion/llm/inference.rb', line 783

def emit_non_pipeline_metering(response, model:, provider:, caller: nil)
  return unless response

  input  = response.respond_to?(:input_tokens)  ? response.input_tokens.to_i  : 0
  output = response.respond_to?(:output_tokens) ? response.output_tokens.to_i : 0
  usage = response.respond_to?(:usage) ? response.usage : {}
  usage_hash = usage.is_a?(Hash) ? usage : {}
  thinking = (usage_hash[:thinking_tokens] || usage_hash[:thinking] || 0).to_i

  finish = nil
  if response.respond_to?(:stop_reason)
    finish = response.stop_reason&.to_s
  elsif response.respond_to?(:stop) && response.stop.is_a?(Hash)
    finish = response.stop[:reason]&.to_s
  end

  meta = response.respond_to?(:meta) ? response.meta : {}
  meta_hash = meta.is_a?(Hash) ? meta : {}
  latency = (meta_hash[:latency_ms] || meta_hash.dig(:timing, :latency_ms) || 0).to_i
  wall_clock = (meta_hash[:wall_clock_ms] || meta_hash.dig(:timing, :wall_clock_ms) || 0).to_i

  Legion::LLM::Metering.emit(
    provider:          provider,
    model_id:          model,
    request_type:      'chat',
    tier:              'direct',
    input_tokens:      input,
    output_tokens:     output,
    thinking_tokens:   thinking,
    total_tokens:      input + output + thinking,
    finish_reason:     finish,
    provider_instance: response.respond_to?(:instance) ? response.instance : nil,
    latency_ms:        latency,
    wall_clock_ms:     wall_clock,
    caller:            caller,
    event_type:        'llm_completion',
    status:            'success'
  )
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.inference.non_pipeline_metering')
end

.emit_privacy_blocked_audit ⇒ Object 



833
834
835
836
837
838
839
840
841
842
 
# File 'lib/legion/llm/inference.rb', line 833

def emit_privacy_blocked_audit
  Legion::LLM::Audit.emit_prompt(
    request_id: nil, conversation_id: nil, caller: Legion::LLM::PublisherIdentity.caller_hash,
    routing: {}, tokens: {}, status: 'privacy_blocked',
    error: { class: 'PrivacyModeError', message: 'External tiers blocked by enterprise privacy' },
    timestamp: Time.now, request_type: 'chat'
  )
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.inference.emit_privacy_blocked_audit')
end

.enterprise_privacy? ⇒ Boolean

Returns:

  • (Boolean) 


825
826
827
828
829
830
831
 
# File 'lib/legion/llm/inference.rb', line 825

def enterprise_privacy?
  if Legion::Settings.respond_to?(:enterprise_privacy?)
    Legion::Settings.enterprise_privacy?
  else
    ENV['LEGION_ENTERPRISE_PRIVACY'] == 'true'
  end
end

.escalation_attempt_passed?(response, result, resolution, duration_ms, history, chain, caller: nil) ⇒ Boolean

Returns:

  • (Boolean) 


624
625
626
627
628
629
630
631
632
633
 
# File 'lib/legion/llm/inference.rb', line 624

def escalation_attempt_passed?(response, result, resolution, duration_ms, history, chain, caller: nil)
  return false unless result.passed

  report_health(:success, resolution, duration_ms)
  history << build_attempt(resolution, :success, [], duration_ms)
  attach_escalation_history(response, history, resolution, chain)
  publish_escalation_event(history, :success, caller: caller) if history.size > 1
  log.debug "[llm][inference] chat_with_escalation success attempts=#{history.size}"
  true
end

.escalation_attempt_response(resolution, message, kwargs) ⇒ Object 



616
617
618
619
620
621
622
 
# File 'lib/legion/llm/inference.rb', line 616

def escalation_attempt_response(resolution, message, kwargs)
  opts = { model: resolution.model, provider: resolution.provider }
  opts.merge!(kwargs.except(*FRAMEWORK_KEYS))
  chat_single_native(model: opts[:model], provider: opts[:provider],
                     message: message, caller: kwargs[:caller],
                     **opts.except(:model, :provider))
end

.escalation_enabled? ⇒ Boolean

Returns:

  • (Boolean) 


767
768
769
770
771
772
773
 
# File 'lib/legion/llm/inference.rb', line 767

def escalation_enabled?
  routing = Legion::Settings[:llm][:routing]
  return false unless routing.is_a?(Hash)

  esc = routing.is_a?(Hash) ? (routing[:escalation] || {}) : {}
  esc[:enabled] == true
end

.escalation_quality_threshold ⇒ Object 



775
776
777
778
779
780
781
 
# File 'lib/legion/llm/inference.rb', line 775

def escalation_quality_threshold
  routing = Legion::Settings[:llm][:routing]
  return 50 unless routing.is_a?(Hash)

  esc = routing.is_a?(Hash) ? (routing[:escalation] || {}) : {}
  esc[:quality_threshold] || 50
end

.external_tier?(tier) ⇒ Boolean

Returns:

  • (Boolean) 


866
867
868
 
# File 'lib/legion/llm/inference.rb', line 866

def external_tier?(tier)
  %i[cloud frontier openai_compat].include?(tier)
end

.extract_error_category_from_attempt(attempt) ⇒ Object 



712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
 
# File 'lib/legion/llm/inference.rb', line 712

def extract_error_category_from_attempt(attempt)
  return nil unless attempt.is_a?(Hash)

  failures = Array(attempt[:failures])
  return nil if failures.empty?

  first = failures.first
  if first.is_a?(Hash)
    first[:category]&.to_s || first[:error]&.to_s
  elsif first.is_a?(String)
    first
  else
    first.to_s
  end
end

.hash_response_details(result, requested_model:, requested_provider:) ⇒ Object 



286
287
288
289
290
291
292
293
294
295
296
297
 
# File 'lib/legion/llm/inference.rb', line 286

def hash_response_details(result, requested_model:, requested_provider:)
  meta = result[:meta] || result['meta'] || {}
  {
    output:        result[:response] || result['response'] || result[:content] || result['content'] || result.dig(:message, :content) || result.to_s,
    provider:      result[:provider] || result['provider'] || meta[:provider] || meta['provider'] || requested_provider,
    model:         result[:model] || result['model'] || meta[:model] || meta['model'] || requested_model,
    input_tokens:  result[:input_tokens] || result['input_tokens'] || meta[:tokens_in] || meta['tokens_in'],
    output_tokens: result[:output_tokens] || result['output_tokens'] || meta[:tokens_out] || meta['tokens_out'],
    stop_reason:   result[:stop_reason] || result['stop_reason'] || result.dig(:stop, :reason) || result.dig('stop', 'reason'),
    tool_calls:    Array(result[:tool_calls] || result['tool_calls'] || result[:tools] || result['tools']).size
  }
end

.inference_input_payload(message:, messages:) ⇒ Object 



241
242
243
244
245
 
# File 'lib/legion/llm/inference.rb', line 241

def inference_input_payload(message:, messages:)
  return messages unless messages.nil?

  message
end

.inference_response_details(result, requested_model:, requested_provider:) ⇒ Object 



263
264
265
266
267
268
269
270
 
# File 'lib/legion/llm/inference.rb', line 263

def inference_response_details(result, requested_model:, requested_provider:)
  if result.is_a?(Legion::LLM::Inference::Response)
    return pipeline_response_details(result, requested_model: requested_model, requested_provider: requested_provider)
  end
  return hash_response_details(result, requested_model: requested_model, requested_provider: requested_provider) if result.is_a?(Hash)

  object_response_details(result, requested_model: requested_model, requested_provider: requested_provider)
end

.inference_text_length(payload) ⇒ Object 



247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
 
# File 'lib/legion/llm/inference.rb', line 247

def inference_text_length(payload)
  case payload
  when Array
    payload.sum { |item| inference_text_length(item) }
  when Hash
    return payload[:content].to_s.length if payload.key?(:content)
    return payload['content'].to_s.length if payload.key?('content')

    payload.values.sum { |item| inference_text_length(item) }
  when nil
    0
  else
    payload.to_s.length
  end
end

.inference_token_value(tokens, key) ⇒ Object 



321
322
323
324
325
326
327
328
329
 
# File 'lib/legion/llm/inference.rb', line 321

def inference_token_value(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, total: :total_tokens }[key]
  return tokens.public_send(method_name) if method_name && tokens.respond_to?(method_name)

  nil
end

.log_inference_error(request_type:, requested_model:, requested_provider:, error:, duration_ms:) ⇒ Object 



221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
 
# File 'lib/legion/llm/inference.rb', line 221

def log_inference_error(request_type:, requested_model:, requested_provider:, error:, duration_ms:)
  parts = [
    '[llm][inference] response',
    "type=#{request_type}",
    'status=error',
    "duration_ms=#{duration_ms}",
    "error_class=#{error.class}",
    "error=#{error.message.inspect}"
  ]
  parts << "requested_provider=#{requested_provider}" if requested_provider
  parts << "requested_model=#{requested_model}" if requested_model
  log.error(parts.join(' '))
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.inference.log_error')
end

.log_inference_request(request_type:, requested_model:, requested_provider:, intent:, tier:, message:, kwargs:) ⇒ Object 



177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
 
# File 'lib/legion/llm/inference.rb', line 177

def log_inference_request(request_type:, requested_model:, requested_provider:, intent:, tier:, message:, kwargs:)
  input = inference_input_payload(message: message, messages: kwargs[:messages])
  parts = [
    '[llm][inference] request',
    "type=#{request_type}",
    "input_length=#{inference_text_length(input)}",
    "input=#{input.inspect}"
  ]
  parts << "requested_provider=#{requested_provider}" if requested_provider
  parts << "requested_model=#{requested_model}" if requested_model
  parts << "intent=#{intent}" if intent
  parts << "tier=#{tier}" if tier
  parts << "caller=#{caller_descriptor(kwargs[:caller])}" if kwargs[:caller]
  parts << "conversation_id=#{kwargs[:conversation_id]}" if kwargs[:conversation_id]
  parts << "request_id=#{kwargs[:request_id]}" if kwargs[:request_id]
  parts << "tools=#{Array(kwargs[:tools]).size}" if kwargs.key?(:tools)
  parts << 'stream=true' if kwargs[:stream]
  log.info(parts.join(' '))
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.inference.log_request')
end

.log_inference_response(request_type:, requested_model:, requested_provider:, result:, duration_ms:) ⇒ Object 



199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
 
# File 'lib/legion/llm/inference.rb', line 199

def log_inference_response(request_type:, requested_model:, requested_provider:, result:, duration_ms:)
  details = inference_response_details(result, requested_model: requested_model, requested_provider: requested_provider)
  parts = [
    '[llm][inference] response',
    "type=#{request_type}",
    'status=ok',
    "duration_ms=#{duration_ms}",
    "result_class=#{result.class}",
    "output_length=#{inference_text_length(details[:output])}",
    "output=#{details[:output].inspect}"
  ]
  parts << "provider=#{details[:provider]}" if details[:provider]
  parts << "model=#{details[:model]}" if details[:model]
  parts << "input_tokens=#{details[:input_tokens]}" unless details[:input_tokens].nil?
  parts << "output_tokens=#{details[:output_tokens]}" unless details[:output_tokens].nil?
  parts << "stop_reason=#{details[:stop_reason]}" if details[:stop_reason]
  parts << "tool_calls=#{details[:tool_calls]}" unless details[:tool_calls].nil?
  log.info(parts.join(' '))
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.inference.log_response')
end

.maybe_shadow_evaluate(response, messages, primary_model) ⇒ Object 



547
548
549
550
551
552
553
554
555
556
557
558
559
560
 
# File 'lib/legion/llm/inference.rb', line 547

def maybe_shadow_evaluate(response, messages, primary_model)
  return unless Quality::ShadowEval.enabled? && Quality::ShadowEval.should_sample?

  log.debug "[llm][inference] shadow_evaluate primary_model=#{primary_model}"
  Inference::Executor::ASYNC_THREAD_POOL.post do
    Quality::ShadowEval.evaluate(
      primary_response: { content: response.respond_to?(:content) ? response.content : response.to_s,
                          model: primary_model, usage: {} },
      messages:         messages
    )
  rescue StandardError => e
    handle_exception(e, level: :warn, operation: 'llm.inference.shadow_eval')
  end
end

.native_provider_error(operation) ⇒ Object 



532
533
534
535
536
 
# File 'lib/legion/llm/inference.rb', line 532

def native_provider_error(operation)
  Legion::LLM::ProviderError.new(
    "Native provider dispatch is required for #{operation}. Configure a registered lex-llm provider."
  )
end

.normalize_ask_direct_hash(result, fallback_model:) ⇒ Object 



466
467
468
469
470
471
472
473
474
475
476
477
478
 
# File 'lib/legion/llm/inference.rb', line 466

def normalize_ask_direct_hash(result, fallback_model:)
  meta = result[:meta].is_a?(Hash) ? result[:meta] : {}
  {
    status:   result[:status] || :done,
    response: result[:response] || result[:content],
    meta:     {
      tier:       meta[:tier] || :direct,
      model:      (meta[:model] || fallback_model).to_s,
      tokens_in:  meta[:tokens_in],
      tokens_out: meta[:tokens_out]
    }
  }
end

.object_response_details(result, requested_model:, requested_provider:) ⇒ Object 



299
300
301
302
303
304
305
306
307
308
309
310
 
# File 'lib/legion/llm/inference.rb', line 299

def object_response_details(result, requested_model:, requested_provider:)
  tool_calls = safe_inference_value(result, :tool_calls)
  {
    output:        safe_inference_value(result, :content) || result.to_s,
    provider:      requested_provider,
    model:         safe_inference_value(result, :model_id)&.to_s || requested_model,
    input_tokens:  safe_inference_value(result, :input_tokens),
    output_tokens: safe_inference_value(result, :output_tokens),
    stop_reason:   safe_inference_value(result, :stop_reason),
    tool_calls:    tool_calls.nil? ? nil : Array(tool_calls).size
  }
end

.pipeline_enabled? ⇒ Boolean

Returns:

  • (Boolean) 


390
391
392
393
394
395
 
# File 'lib/legion/llm/inference.rb', line 390

def pipeline_enabled?
  Legion::Settings[:llm][:pipeline_enabled] == true
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.inference.pipeline_enabled')
  false
end

.pipeline_response_details(result, requested_model:, requested_provider:) ⇒ Object 



272
273
274
275
276
277
278
279
280
281
282
283
284
 
# File 'lib/legion/llm/inference.rb', line 272

def pipeline_response_details(result, requested_model:, requested_provider:)
  message = result.message
  tokens = result.tokens
  {
    output:        message.is_a?(Hash) ? (message[:content] || message['content']) : message.to_s,
    provider:      result.routing[:provider] || result.routing['provider'] || requested_provider,
    model:         result.routing[:model] || result.routing['model'] || requested_model,
    input_tokens:  inference_token_value(tokens, :input),
    output_tokens: inference_token_value(tokens, :output),
    stop_reason:   result.stop[:reason] || result.stop['reason'],
    tool_calls:    Array(result.tools).size
  }
end

.publish_escalation_event(history, final_outcome, caller: nil) ⇒ Object 



677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
 
# File 'lib/legion/llm/inference.rb', line 677

def publish_escalation_event(history, final_outcome, caller: nil)
  return if history.size <= 1

  first_attempt = history.first
  last_attempt = history.last

  event = {
    outcome:        final_outcome,
    attempts:       history.size,
    history:        history,
    # Flat fields for consumer compatibility
    from_provider:  first_attempt&.dig(:provider),
    from_instance:  first_attempt&.dig(:instance),
    from_model:     first_attempt&.dig(:model),
    to_provider:    last_attempt&.dig(:provider),
    to_instance:    last_attempt&.dig(:instance),
    to_model:       last_attempt&.dig(:model),
    reason:         ('provider_failover' if [:failure, 'failure', :error].include?(first_attempt&.dig(:outcome))),
    error_category: extract_error_category_from_attempt(first_attempt),
    attempt_no:     history.size,
    latency_ms:     history.sum { |a| (a[:duration_ms] || 0).to_i },
    caller:         caller || Legion::LLM::PublisherIdentity.caller_hash,
    timestamp:      Time.now.utc.iso8601
  }

  Legion::Events.emit('llm.escalation', **event) if defined?(Legion::Events) && Legion::Events.respond_to?(:emit)

  log.info "[llm][inference] escalation_event outcome=#{final_outcome} attempts=#{history.size}"

  Transport::Messages::EscalationEvent.new(event).publish if Legion::Settings.dig(:transport, :connected) == true
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.inference.publish_escalation_event', outcome: final_outcome)
  nil
end

.record_escalation_error(error, resolution, duration_ms, history) ⇒ Object 



635
636
637
638
639
640
641
642
643
644
645
646
647
 
# File 'lib/legion/llm/inference.rb', line 635

def record_escalation_error(error, resolution, duration_ms, history)
  handle_exception(
    error,
    level:     :warn,
    handled:   true,
    operation: 'llm.inference.escalation_attempt',
    model:     resolution&.model,
    provider:  resolution&.provider,
    tier:      resolution&.tier
  )
  report_health(:error, resolution, duration_ms) if resolution
  history << build_attempt(resolution, :error, [error.class.name], duration_ms) if resolution
end

.report_health(signal, resolution, duration_ms, failures: nil) ⇒ Object 



666
667
668
669
670
671
672
673
674
675
 
# File 'lib/legion/llm/inference.rb', line 666

def report_health(signal, resolution, duration_ms, failures: nil)
  return unless Router.routing_enabled?

  metadata = { duration_ms: duration_ms }
  metadata[:failures] = failures if failures
  Router.health_tracker.report(provider: resolution.provider, offering_id: resolution.offering_id,
                               signal: signal, value: 1, metadata: metadata)
  Router.health_tracker.report(provider: resolution.provider, offering_id: resolution.offering_id,
                               signal: :latency, value: duration_ms, metadata: {})
end

.resolve_ask_direct_response(result, message, requested_model) ⇒ Object 



452
453
454
455
456
457
458
459
460
461
462
463
464
 
# File 'lib/legion/llm/inference.rb', line 452

def resolve_ask_direct_response(result, message, requested_model, &)
  if direct_chat_session?(result)
    response = block_given? ? result.ask(message, &) : result.ask(message)
    return [response, result.model.to_s]
  end

  resolved_model = if result.respond_to?(:model_id) && result.model_id
                     result.model_id.to_s
                   else
                     (requested_model || Legion::Settings[:llm][:default_model]).to_s
                   end
  [result, resolved_model]
end

.response_guards_enabled? ⇒ Boolean

Returns:

  • (Boolean) 


728
729
730
 
# File 'lib/legion/llm/inference.rb', line 728

def response_guards_enabled?
  Legion::Settings.dig(:llm, :response_guards, :enabled) == true
end

.run_escalation_attempt(resolution, message:, kwargs:, threshold:, quality_check:, history:, chain:) ⇒ Object 



593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
 
# File 'lib/legion/llm/inference.rb', line 593

def run_escalation_attempt(resolution, message:, kwargs:, threshold:, quality_check:, history:, chain:)
  start_time = Time.now
  assert_external_allowed! if resolution.respond_to?(:external?) && resolution.external?

  response = escalation_attempt_response(resolution, message, kwargs)
  duration_ms = ((Time.now - start_time) * 1000).round
  result = Quality::Checker.check(response, quality_threshold: threshold, quality_check: quality_check)

  return [response, nil] if escalation_attempt_passed?(response, result, resolution, duration_ms, history, chain,
                                                       caller: kwargs[:caller])

  report_health(:quality_failure, resolution, duration_ms, failures: result.failures)
  history << build_attempt(resolution, :quality_failure, result.failures, duration_ms)
  log.debug "[llm][inference] chat_with_escalation quality_failure attempt=#{history.size} failures=#{result.failures}"
  [nil, nil]
rescue Legion::LLM::PrivacyModeError
  raise
rescue StandardError => e
  duration_ms = ((Time.now - start_time) * 1000).round
  record_escalation_error(e, resolution, duration_ms, history)
  [nil, e]
end

.safe_inference_value(object, method_name) ⇒ Object 



312
313
314
315
316
317
318
319
 
# File 'lib/legion/llm/inference.rb', line 312

def safe_inference_value(object, method_name)
  return nil unless object.methods.include?(method_name) || object.private_methods.include?(method_name)

  object.public_send(method_name)
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.inference.safe_value', method_name: method_name)
  nil
end

.try_defer(intent:, urgency:, model:, provider:, message:) ⇒ Object 



538
539
540
541
542
543
544
545
 
# File 'lib/legion/llm/inference.rb', line 538

def try_defer(intent:, urgency:, model:, provider:, message:, **)
  return nil unless Scheduling.enabled? && Scheduling.should_defer?(intent: intent || :normal, urgency: urgency)
  return nil unless Legion::LLM::Scheduling::Batch.enabled?

  log.debug "[llm][inference] try_defer deferring intent=#{intent} urgency=#{urgency}"
  entry_id = Legion::LLM::Scheduling::Batch.enqueue(model: model, provider: provider, message: message, priority: urgency, **)
  { deferred: true, batch_id: entry_id, next_off_peak: Scheduling.next_off_peak.iso8601 }
end