Module: Legion::LLM

Extended by:
Providers, Legion::Logging::Helper
Defined in:
lib/legion/llm.rb,
lib/legion/llm/audit.rb,
lib/legion/llm/batch.rb,
lib/legion/llm/cache.rb,
lib/legion/llm/fleet.rb,
lib/legion/llm/hooks.rb,
lib/legion/llm/usage.rb,
lib/legion/llm/errors.rb,
lib/legion/llm/helper.rb,
lib/legion/llm/router.rb,
lib/legion/llm/routes.rb,
lib/legion/llm/skills.rb,
lib/legion/llm/version.rb,
lib/legion/llm/metering.rb,
lib/legion/llm/off_peak.rb,
lib/legion/llm/pipeline.rb,
lib/legion/llm/settings.rb,
lib/legion/llm/arbitrage.rb,
lib/legion/llm/providers.rb,
lib/legion/llm/compressor.rb,
lib/legion/llm/embeddings.rb,
lib/legion/llm/scheduling.rb,
lib/legion/llm/fleet/error.rb,
lib/legion/llm/router/rule.rb,
lib/legion/llm/shadow_eval.rb,
lib/legion/llm/skills/base.rb,
lib/legion/llm/cost_tracker.rb,
lib/legion/llm/daemon_client.rb,
lib/legion/llm/fleet/handler.rb,
lib/legion/llm/fleet/request.rb,
lib/legion/llm/skills/errors.rb,
lib/legion/llm/token_tracker.rb,
lib/legion/llm/tools/adapter.rb,
lib/legion/llm/audit/exchange.rb,
lib/legion/llm/cost_estimator.rb,
lib/legion/llm/fleet/exchange.rb,
lib/legion/llm/fleet/response.rb,
lib/legion/llm/hooks/metering.rb,
lib/legion/llm/metering/event.rb,
lib/legion/llm/pipeline/steps.rb,
lib/legion/llm/response_cache.rb,
lib/legion/llm/context_curator.rb,
lib/legion/llm/hooks/rag_guard.rb,
lib/legion/llm/native_dispatch.rb,
lib/legion/llm/quality_checker.rb,
lib/legion/llm/skills/registry.rb,
lib/legion/llm/skills/settings.rb,
lib/legion/llm/audit/tool_event.rb,
lib/legion/llm/confidence_score.rb,
lib/legion/llm/discovery/ollama.rb,
lib/legion/llm/discovery/system.rb,
lib/legion/llm/fleet/dispatcher.rb,
lib/legion/llm/hooks/reflection.rb,
lib/legion/llm/pipeline/profile.rb,
lib/legion/llm/pipeline/request.rb,
lib/legion/llm/pipeline/tracing.rb,
lib/legion/llm/tools/dispatcher.rb,
lib/legion/llm/audit/skill_event.rb,
lib/legion/llm/confidence_scorer.rb,
lib/legion/llm/hooks/reciprocity.rb,
lib/legion/llm/metering/exchange.rb,
lib/legion/llm/pipeline/executor.rb,
lib/legion/llm/pipeline/response.rb,
lib/legion/llm/pipeline/timeline.rb,
lib/legion/llm/provider_registry.rb,
lib/legion/llm/router/resolution.rb,
lib/legion/llm/structured_output.rb,
lib/legion/llm/tools/interceptor.rb,
lib/legion/llm/transport/message.rb,
lib/legion/llm/audit/prompt_event.rb,
lib/legion/llm/conversation_store.rb,
lib/legion/llm/escalation_history.rb,
lib/legion/llm/escalation_tracker.rb,
lib/legion/llm/hooks/budget_guard.rb,
lib/legion/llm/skills/disk_loader.rb,
lib/legion/llm/skills/step_result.rb,
lib/legion/llm/codex_config_loader.rb,
lib/legion/llm/hooks/cost_tracking.rb,
lib/legion/llm/override_confidence.rb,
lib/legion/llm/pipeline/steps/rbac.rb,
lib/legion/llm/claude_config_loader.rb,
lib/legion/llm/hooks/response_guard.rb,
lib/legion/llm/pipeline/gaia_caller.rb,
lib/legion/llm/pipeline/steps/debate.rb,
lib/legion/llm/pipeline/tool_adapter.rb,
lib/legion/llm/router/health_tracker.rb,
lib/legion/llm/fleet/reply_dispatcher.rb,
lib/legion/llm/pipeline/steps/billing.rb,
lib/legion/llm/pipeline/steps/metering.rb,
lib/legion/llm/router/escalation_chain.rb,
lib/legion/llm/skills/skill_run_result.rb,
lib/legion/llm/pipeline/audit_publisher.rb,
lib/legion/llm/pipeline/steps/rag_guard.rb,
lib/legion/llm/pipeline/tool_dispatcher.rb,
lib/legion/llm/pipeline/steps/tool_calls.rb,
lib/legion/llm/skills/external_discovery.rb,
lib/legion/llm/transport/exchanges/audit.rb,
lib/legion/llm/pipeline/steps/rag_context.rb,
lib/legion/llm/router/gateway_interceptor.rb,
lib/legion/llm/pipeline/steps/prompt_cache.rb,
lib/legion/llm/pipeline/steps/token_budget.rb,
lib/legion/llm/pipeline/enrichment_injector.rb,
lib/legion/llm/pipeline/steps/gaia_advisory.rb,
lib/legion/llm/pipeline/steps/mcp_discovery.rb,
lib/legion/llm/pipeline/steps/post_response.rb,
lib/legion/llm/pipeline/steps/tier_assigner.rb,
lib/legion/llm/pipeline/steps/trigger_match.rb,
lib/legion/llm/transport/exchanges/metering.rb,
lib/legion/llm/pipeline/steps/classification.rb,
lib/legion/llm/pipeline/steps/skill_injector.rb,
lib/legion/llm/pipeline/steps/span_annotator.rb,
lib/legion/llm/pipeline/steps/tool_discovery.rb,
lib/legion/llm/tools/interceptors/python_venv.rb,
lib/legion/llm/transport/exchanges/escalation.rb,
lib/legion/llm/transport/messages/audit_event.rb,
lib/legion/llm/patches/ruby_llm_parallel_tools.rb,
lib/legion/llm/pipeline/steps/knowledge_capture.rb,
lib/legion/llm/pipeline/steps/confidence_scoring.rb,
lib/legion/llm/transport/messages/escalation_event.rb

Defined Under Namespace

Modules: Arbitrage, Audit, Batch, Cache, ClaudeConfigLoader, CodexConfigLoader, Compressor, ConfidenceScorer, ConversationStore, CostEstimator, CostTracker, DaemonClient, Discovery, Embeddings, EscalationHistory, EscalationTracker, Fleet, Helper, Hooks, Metering, NativeDispatch, OffPeak, OverrideConfidence, Patches, Pipeline, ProviderRegistry, Providers, QualityChecker, ResponseCache, Router, Routes, Scheduling, Settings, ShadowEval, Skills, StructuredOutput, TokenTracker, Tools, Transport Classes: AuthError, ConfidenceScore, ContextCurator, ContextOverflow, DaemonDeniedError, DaemonRateLimitedError, EmbeddingUnavailableError, EscalationExhausted, LLMError, NativeResponseAdapter, PipelineError, PrivacyModeError, ProviderDown, ProviderError, RateLimitError, TokenBudgetExceeded, UnsupportedCapability, Usage

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
VERSION = 
'0.7.3'

Constants included from Providers

Providers::SAAS_PROVIDERS

Class Attribute Summary collapse

Class Method Summary collapse

Methods included from Providers

apply_provider_config, auto_enable_from_resolved_credentials, broker_has_credential?, configure_anthropic, configure_azure, configure_bedrock, configure_gemini, configure_ollama, configure_openai, configure_providers, ollama_running?, recover_openai_with_codex, recover_with_alternative_credentials, resolve_broker_aws_credentials, resolve_broker_credential, verify_providers, verify_single_provider

Class Attribute Details

.embedding_fallback_chain ⇒ Object (readonly)

Returns the value of attribute embedding_fallback_chain.



109
110
111
 
# File 'lib/legion/llm.rb', line 109

def embedding_fallback_chain
  @embedding_fallback_chain
end

.embedding_model ⇒ Object (readonly)

Returns the value of attribute embedding_model.



109
110
111
 
# File 'lib/legion/llm.rb', line 109

def embedding_model
  @embedding_model
end

.embedding_provider ⇒ Object (readonly)

Returns the value of attribute embedding_provider.



109
110
111
 
# File 'lib/legion/llm.rb', line 109

def embedding_provider
  @embedding_provider
end

Class Method Details

.agent(agent_class) ⇒ RubyLLM::Agent

Create a configured agent instance

Parameters:

  • agent_class (Class) —

    a RubyLLM::Agent subclass

  • kwargs (Hash) —

    additional options

Returns:

  • (RubyLLM::Agent) 


311
312
313
 
# File 'lib/legion/llm.rb', line 311

def agent(agent_class, **)
  agent_class.new(**)
end

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

Send a single message — daemon-first, falls through to direct on unavailability.



167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
 
# File 'lib/legion/llm.rb', line 167

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 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

.can_embed? ⇒ Boolean

Returns:

  • (Boolean) 


105
106
107
 
# File 'lib/legion/llm.rb', line 105

def can_embed?
  @can_embed == true
end

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

Create a new chat session — delegates to lex-llm-gateway when available for automatic metering and fleet dispatch



121
122
123
124
125
126
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
 
# File 'lib/legion/llm.rb', line 121

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 || 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

Direct chat bypassing gateway — used by gateway runners to avoid recursion



217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
 
# File 'lib/legion/llm.rb', line 217

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] 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 'Legion::LLM 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] 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] chat_direct.exit result_class=#{result.class} result_nil=#{result.nil?}")

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

  result
end

.embed(text) ⇒ Object

Generate embeddings



268
269
270
271
272
273
274
275
276
 
# File 'lib/legion/llm.rb', line 268

def embed(text, **)
  if defined?(Legion::Telemetry::OpenInference)
    Legion::Telemetry::OpenInference.embedding_span(
      model: (settings[:default_model] || 'unknown').to_s
    ) { |_span| embed_direct(text, **) }
  else
    embed_direct(text, **)
  end
end

.embed_batch(texts) ⇒ Array<Hash>

Batch embed multiple texts

Parameters:

  • texts (Array<String>) —

    texts to embed

Returns:

  • (Array<Hash>) 


286
287
288
 
# File 'lib/legion/llm.rb', line 286

def embed_batch(texts, **)
  Embeddings.generate_batch(texts: texts, **)
end

.embed_direct(text) ⇒ Object

Direct embed bypassing gateway



279
280
281
 
# File 'lib/legion/llm.rb', line 279

def embed_direct(text, **)
  Embeddings.generate(text: text, **)
end

.settings ⇒ Object 



111
112
113
114
115
116
117
 
# File 'lib/legion/llm.rb', line 111

def settings
  if Legion.const_defined?('Settings', false)
    Legion::Settings[:llm]
  else
    Legion::LLM::Settings.default
  end
end

.shutdown ⇒ Object 



90
91
92
93
94
95
96
97
98
99
 
# File 'lib/legion/llm.rb', line 90

def shutdown
  Legion::Settings[:llm][:connected] = false
  @started = false
  @can_embed = nil
  @embedding_provider = nil
  @embedding_model = nil
  @embedding_fallback_chain = nil
  ProviderRegistry.reset!
  log.info 'Legion::LLM shut down'
end

.start ⇒ Object 



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
 
# File 'lib/legion/llm.rb', line 61

def start
  log.debug 'Legion::LLM is running start'

  require 'legion/llm/claude_config_loader'
  ClaudeConfigLoader.load

  require 'legion/llm/codex_config_loader'
  CodexConfigLoader.load

  resolve_llm_secrets
  configure_providers
  verify_providers
  run_discovery
  detect_embedding_capability
  set_defaults
  auto_register_providers

  install_hooks
  load_tool_interceptors

  # Skills startup — load after settings, before pipeline is used
  Legion::LLM::Skills.start if defined?(Legion::LLM::Skills) && settings.dig(:skills, :enabled) != false

  @started = true
  Legion::Settings[:llm][:connected] = true
  log.info 'Legion::LLM started'
  register_routes
end

.started? ⇒ Boolean

Returns:

  • (Boolean) 


101
102
103
 
# File 'lib/legion/llm.rb', line 101

def started?
  @started == true
end

.structured(messages:, schema:) ⇒ Object

Generate structured JSON output



291
292
293
294
295
296
297
298
299
 
# File 'lib/legion/llm.rb', line 291

def structured(messages:, schema:, **)
  if defined?(Legion::Telemetry::OpenInference)
    Legion::Telemetry::OpenInference.llm_span(
      model: (settings[:default_model] || 'unknown').to_s, input: messages.to_s
    ) { |_span| structured_direct(messages: messages, schema: schema, **) }
  else
    structured_direct(messages: messages, schema: schema, **)
  end
end

.structured_direct(messages:, schema:) ⇒ Object

Direct structured bypassing gateway



302
303
304
305
 
# File 'lib/legion/llm.rb', line 302

def structured_direct(messages:, schema:, **)
  require 'legion/llm/structured_output'
  StructuredOutput.generate(messages: messages, schema: schema, **)
end