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/tool_adapter.rb,
lib/legion/llm/inference/steps/billing.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/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, Profile, Prompt, 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
ToolAdapter = 
Tools::Adapter
McpToolAdapter = 
Tools::Adapter
ToolDispatcher = 
Tools::Dispatcher

Class Method Summary collapse

Class Method Details

.adapted_registry_tools ⇒ Object 



526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
 
# File 'lib/legion/llm/inference.rb', line 526

def adapted_registry_tools
  tool_classes = if defined?(::Legion::Tools::Registry)
                   ::Legion::Tools::Registry.tools
                 else
                   return []
                 end

  tool_classes.map do |tool_class|
    ToolAdapter.new(tool_class)
  rescue StandardError => e
    handle_exception(e, level: :warn, operation: 'llm.inference.adapted_registry_tools', tool_class: tool_class.to_s)
    nil
  end.compact
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.inference.adapted_registry_tools')
  []
end

.apply_response_guards(result, kwargs) ⇒ Object 



665
666
667
668
669
670
671
672
673
674
675
676
677
678
 
# File 'lib/legion/llm/inference.rb', line 665

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 



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
 
# File 'lib/legion/llm/inference.rb', line 79

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 



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
447
 
# File 'lib/legion/llm/inference.rb', line 422

def ask_direct(message:, model: nil, provider: nil, intent: nil, tier: nil, &)
  assert_cloud_allowed! if effective_tier_is_cloud?(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::LLM.settings[: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_cloud_allowed! ⇒ Object

Raises:



718
719
720
721
722
723
724
 
# File 'lib/legion/llm/inference.rb', line 718

def assert_cloud_allowed!
  return unless enterprise_privacy?

  raise Legion::LLM::PrivacyModeError,
        'Cloud LLM tier is disabled: enterprise_data_privacy is enabled. ' \
        'Only Tier 0 (cache) and Tier 1 (local Ollama) are permitted.'
end

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



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

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

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



620
621
622
623
 
# File 'lib/legion/llm/inference.rb', line 620

def build_attempt(resolution, outcome, failures, duration_ms)
  { model: resolution.model, provider: resolution.provider, tier: resolution.tier,
    outcome: outcome, failures: failures, duration_ms: duration_ms }
end

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



684
685
686
687
688
689
690
691
692
 
# File 'lib/legion/llm/inference.rb', line 684

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::LLM.settings[:default_model],
    provider:    provider || Legion::LLM.settings[:default_provider],
    messages:    messages_arr,
    temperature: temperature
  )
end

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

Returns:

  • (Boolean) 


680
681
682
 
# File 'lib/legion/llm/inference.rb', line 680

def cacheable?(cache_opt, temperature, message)
  cache_opt != false && temperature.to_f.zero? && message && Cache.enabled?
end

.caller_descriptor(caller_context) ⇒ Object 



332
333
334
335
336
337
338
339
340
341
 
# File 'lib/legion/llm/inference.rb', line 332

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



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
 
# File 'lib/legion/llm/inference.rb', line 32

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::LLM.settings[: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 



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
176
 
# File 'lib/legion/llm/inference.rb', line 128

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::LLM.settings.dig(: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, &block) ⇒ Object 



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
514
515
516
517
518
519
520
521
522
523
524
 
# File 'lib/legion/llm/inference.rb', line 481

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

  if (intent || tier) && Router.routing_enabled?
    resolution = Router.resolve(intent: intent, tier: tier, model: model, provider: provider)
    if resolution
      resolution = Router::GatewayInterceptor.intercept(resolution, context: kwargs.fetch(:context, {}))
      model = resolution.model
      provider = resolution.provider
      assert_cloud_allowed! if resolution.tier.to_sym == :cloud
    end
  elsif tier
    assert_cloud_allowed! if tier.to_sym == :cloud
  end

  model ||= Legion::LLM.settings[:default_model]
  provider ||= Legion::LLM.settings[: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)

  log.debug "[llm][inference] chat_single model=#{opts[:model]} provider=#{opts[:provider]} message_present=#{!message.nil?} tools=#{tools&.size || 0}"
  session = RubyLLM.chat(**opts)
  tools&.each { |tool| session.with_tool(tool) }
  return session unless message

  log.debug '[llm][inference] chat_single asking session'
  response = block ? session.ask(message, &block) : session.ask(message)
  log.debug "[llm][inference] chat_single response_class=#{response.class} response_nil=#{response.nil?}"

  if response && !block && defined?(Quality::ShadowEval) && Quality::ShadowEval.enabled?
    msgs = session.respond_to?(:messages) ? session.messages : nil
    maybe_shadow_evaluate(response, msgs, opts[:model])
  end

  response
end

.chat_via_pipeline(&block) ⇒ Object 



398
399
400
401
402
 
# File 'lib/legion/llm/inference.rb', line 398

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:



568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
 
# File 'lib/legion/llm/inference.rb', line 568

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 = []

  chain.each do |resolution|
    start_time = Time.now
    begin
      opts = { model: resolution.model, provider: resolution.provider }
      opts.merge!(kwargs.except(*FRAMEWORK_KEYS))
      chat_obj = RubyLLM.chat(**opts)
      response = chat_obj.ask(message)

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

      if 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) if history.size > 1
        log.debug "[llm][inference] chat_with_escalation success attempts=#{history.size}"
        return response
      else
        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}"
      end
    rescue StandardError => e
      duration_ms = ((Time.now - start_time) * 1000).round
      handle_exception(
        e,
        level:     :warn,
        operation: 'llm.inference.escalation_attempt',
        model:     resolution.model,
        provider:  resolution.provider,
        tier:      resolution.tier
      )
      report_health(:error, resolution, duration_ms)
      history << build_attempt(resolution, :error, [e.class.name], duration_ms)
    end
  end

  publish_escalation_event(history, :exhausted) if history.size > 1
  raise Legion::LLM::EscalationExhausted, "All #{history.size} escalation attempts failed"
end

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

rubocop:disable Lint/UnusedMethodArgument



404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
 
# File 'lib/legion/llm/inference.rb', line 404

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) 


449
450
451
 
# File 'lib/legion/llm/inference.rb', line 449

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 



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
389
 
# File 'lib/legion/llm/inference.rb', line 343

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::LLM.settings[:default_model]

  if defined?(Legion::LLM::Hooks)
    blocked = Legion::LLM::Hooks.run_before(messages: messages, model: resolved_model)
    return blocked[:response] 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] 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_cloud?(tier, provider) ⇒ Boolean

Returns:

  • (Boolean) 


726
727
728
729
730
731
732
733
 
# File 'lib/legion/llm/inference.rb', line 726

def effective_tier_is_cloud?(tier, provider)
  return tier.to_sym == :cloud if tier
  return false unless enterprise_privacy?

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

.elapsed_ms_since(started_at) ⇒ Object 



238
239
240
 
# File 'lib/legion/llm/inference.rb', line 238

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

.enterprise_privacy? ⇒ Boolean

Returns:

  • (Boolean) 


710
711
712
713
714
715
716
 
# File 'lib/legion/llm/inference.rb', line 710

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

.escalation_enabled? ⇒ Boolean

Returns:

  • (Boolean) 


694
695
696
697
698
699
700
 
# File 'lib/legion/llm/inference.rb', line 694

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

  esc = routing[:escalation] || {}
  esc[:enabled] == true
end

.escalation_quality_threshold ⇒ Object 



702
703
704
705
706
707
708
 
# File 'lib/legion/llm/inference.rb', line 702

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

  esc = routing[:escalation] || {}
  esc.fetch(:quality_threshold, 50)
end

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



287
288
289
290
291
292
293
294
295
296
297
298
 
# File 'lib/legion/llm/inference.rb', line 287

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 



242
243
244
245
246
 
# File 'lib/legion/llm/inference.rb', line 242

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

  message
end

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



264
265
266
267
268
269
270
271
 
# File 'lib/legion/llm/inference.rb', line 264

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 



248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
 
# File 'lib/legion/llm/inference.rb', line 248

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 



322
323
324
325
326
327
328
329
330
 
# File 'lib/legion/llm/inference.rb', line 322

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 



222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
 
# File 'lib/legion/llm/inference.rb', line 222

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: :debug, operation: 'llm.inference.log_error')
end

.log_inference_request(request_type:, requested_model:, requested_provider:, intent:, tier:, message:, kwargs:) ⇒ 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/inference.rb', line 178

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: :debug, operation: 'llm.inference.log_request')
end

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



200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
 
# File 'lib/legion/llm/inference.rb', line 200

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: :debug, operation: 'llm.inference.log_response')
end

.maybe_shadow_evaluate(response, messages, primary_model) ⇒ Object 



553
554
555
556
557
558
559
560
561
562
563
564
565
566
 
# File 'lib/legion/llm/inference.rb', line 553

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}"
  Thread.new 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: :debug, operation: 'llm.inference.shadow_eval')
  end
end

.normalize_ask_direct_hash(result, fallback_model:) ⇒ Object 



467
468
469
470
471
472
473
474
475
476
477
478
479
 
# File 'lib/legion/llm/inference.rb', line 467

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 



300
301
302
303
304
305
306
307
308
309
310
311
 
# File 'lib/legion/llm/inference.rb', line 300

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) 


391
392
393
394
395
396
 
# File 'lib/legion/llm/inference.rb', line 391

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

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



273
274
275
276
277
278
279
280
281
282
283
284
285
 
# File 'lib/legion/llm/inference.rb', line 273

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) ⇒ Object 



643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
 
# File 'lib/legion/llm/inference.rb', line 643

def publish_escalation_event(history, final_outcome)
  payload = {
    outcome:   final_outcome,
    attempts:  history.size,
    history:   history,
    timestamp: Time.now.utc.iso8601
  }

  Legion::Events.emit('llm.escalation', **payload) 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(payload).publish if defined?(Legion::Settings) && Legion::Settings[:transport][:connected] == true
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.inference.publish_escalation_event', outcome: final_outcome)
  nil
end

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



634
635
636
637
638
639
640
641
 
# File 'lib/legion/llm/inference.rb', line 634

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, signal: signal, value: 1, metadata: metadata)
  Router.health_tracker.report(provider: resolution.provider, signal: :latency, value: duration_ms, metadata: {})
end

.resolve_ask_direct_response(result, message, requested_model) ⇒ Object 



453
454
455
456
457
458
459
460
461
462
463
464
465
 
# File 'lib/legion/llm/inference.rb', line 453

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::LLM.settings[:default_model]).to_s
                   end
  [result, resolved_model]
end

.response_guards_enabled? ⇒ Boolean

Returns:

  • (Boolean) 


661
662
663
 
# File 'lib/legion/llm/inference.rb', line 661

def response_guards_enabled?
  Legion::LLM.settings.dig(:response_guards, :enabled) == true
end

.safe_inference_value(object, method_name) ⇒ Object 



313
314
315
316
317
318
319
320
 
# File 'lib/legion/llm/inference.rb', line 313

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: :debug, operation: 'llm.inference.safe_value', method_name: method_name)
  nil
end

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



544
545
546
547
548
549
550
551
 
# File 'lib/legion/llm/inference.rb', line 544

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