Module: Legion::Extensions::Apollo::Runners::Gas

Defined in:
lib/legion/extensions/apollo/runners/gas.rb

Constant Summary collapse

RELATION_TYPES = 
%w[
  similar_to contradicts depends_on causes
  part_of supersedes supports_by extends
].freeze
RELATE_CONFIDENCE_GATE = 
0.7
SYNTHESIS_CONFIDENCE_CAP = 
0.7
MAX_ANTICIPATIONS = 
3

Class Method Summary collapse

Class Method Details

.build_synthesis_entry(item, facts) ⇒ Object 



249
250
251
252
253
254
255
256
257
258
259
260
261
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 249

def build_synthesis_entry(item, facts)
  source_indices = item[:source_indices] || item['source_indices'] || []
  source_confs = source_indices.filter_map { |i| facts[i]&.dig(:confidence) }
  geo_mean = source_confs.empty? ? 0.5 : geometric_mean(source_confs)

  {
    content:        item[:content] || item['content'],
    content_type:   (item[:content_type] || item['content_type'] || 'inference').to_sym,
    status:         :candidate,
    confidence:     [geo_mean, SYNTHESIS_CONFIDENCE_CAP].min,
    source_indices: source_indices
  }
end

.classify_relation(fact, entry) ⇒ Object 



141
142
143
144
145
146
147
148
149
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 141

def classify_relation(fact, entry)
  if llm_available?
    llm_classify_relation(fact, entry)
  else
    { from_content: fact[:content], to_id: entry[:id], relation_type: 'similar_to', confidence: 0.5 }
  end
rescue StandardError
  { from_content: fact[:content], to_id: entry[:id], relation_type: 'similar_to', confidence: 0.5 }
end

.fallback_relation(fact, entry) ⇒ Object 



200
201
202
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 200

def fallback_relation(fact, entry)
  { from_content: fact[:content], to_id: entry[:id], relation_type: 'similar_to', confidence: 0.5 }
end

.fetch_similar_entries(facts) ⇒ Object 



130
131
132
133
134
135
136
137
138
139
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 130

def fetch_similar_entries(facts)
  entries = []
  facts.each do |fact|
    result = Runners::Knowledge.retrieve_relevant(query: fact[:content], limit: 3, min_confidence: 0.3)
    entries.concat(result[:entries]) if result[:success] && result[:entries]&.any?
  rescue StandardError
    next
  end
  entries.uniq { |e| e[:id] }
end

.geometric_mean(values) ⇒ Object 



263
264
265
266
267
268
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 263

def geometric_mean(values)
  return 0.0 if values.empty?

  product = values.reduce(1.0) { |acc, v| acc * v }
  product**(1.0 / values.length)
end

.llm_anticipate(facts) ⇒ Object 



270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 270

def llm_anticipate(facts)
  facts_text = facts.map { |f| "(#{f[:content_type]}) #{f[:content]}" }.join("\n")

  prompt = <<~PROMPT
    Given these knowledge entries, generate 1-3 likely follow-up questions a user might ask.

    Knowledge:
    #{facts_text}

    Return JSON with a "questions" array of question strings.
  PROMPT

  result = Legion::LLM::Pipeline::GaiaCaller.structured(
    message: prompt.strip,
    schema:  {
      type:       :object,
      properties: {
        questions: { type: :array, items: { type: :string } }
      },
      required:   ['questions']
    },
    phase:   'gas_anticipate'
  )

  content = result.respond_to?(:message) ? result.message[:content] : result.to_s
  parsed = Legion::JSON.load(content)
  questions = parsed.is_a?(Hash) ? (parsed[:questions] || parsed['questions'] || []) : []
  questions = questions.first(MAX_ANTICIPATIONS)

  questions.map do |q|
    promote_to_pattern_store(question: q, facts: facts)
    { question: q }
  end
rescue StandardError
  []
end

.llm_available? ⇒ Boolean

Returns:

  • (Boolean) 


319
320
321
322
323
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 319

def llm_available?
  defined?(Legion::LLM::Pipeline::GaiaCaller)
rescue StandardError
  false
end

.llm_classify_relation(fact, entry) ⇒ Object 



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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 151

def llm_classify_relation(fact, entry)
  prompt = <<~PROMPT
    Classify the relationship between these two knowledge entries.
    Valid types: #{RELATION_TYPES.join(', ')}

    Entry A (new): #{fact[:content]}
    Entry B (existing): #{entry[:content]}

    Return JSON with relation_type and confidence (0.0-1.0).
  PROMPT

  result = Legion::LLM::Pipeline::GaiaCaller.structured(
    message: prompt.strip,
    schema:  {
      type:       :object,
      properties: {
        relations: {
          type:  :array,
          items: {
            type:       :object,
            properties: {
              relation_type: { type: :string },
              confidence:    { type: :number }
            },
            required:   %w[relation_type confidence]
          }
        }
      },
      required:   ['relations']
    },
    phase:   'gas_relate'
  )

  content = result.respond_to?(:message) ? result.message[:content] : result.to_s
  parsed = Legion::JSON.load(content)
  rels = parsed.is_a?(Hash) ? (parsed[:relations] || parsed['relations'] || []) : []
  best = rels.max_by { |r| r[:confidence] || r['confidence'] || 0 }

  return fallback_relation(fact, entry) unless best

  conf = best[:confidence] || best['confidence'] || 0
  rtype = best[:relation_type] || best['relation_type']
  return fallback_relation(fact, entry) if conf < RELATE_CONFIDENCE_GATE || !RELATION_TYPES.include?(rtype)

  { from_content: fact[:content], to_id: entry[:id], relation_type: rtype, confidence: conf }
rescue StandardError
  fallback_relation(fact, entry)
end

.llm_comprehend(messages, response) ⇒ Object 



329
330
331
332
333
334
335
336
337
338
339
340
341
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
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 329

def llm_comprehend(messages, response)
  prompt = <<~PROMPT
    Extract distinct facts from this exchange. Return JSON array of {content:, content_type:} where content_type is one of: fact, concept, procedure, association.

    User: #{messages.last&.dig(:content)}
    Assistant: #{response}
  PROMPT

  result = Legion::LLM::Pipeline::GaiaCaller.structured(
    message: prompt.strip,
    schema:  {
      type:       :object,
      properties: {
        facts: {
          type:  :array,
          items: {
            type:       :object,
            properties: {
              content:      { type: :string },
              content_type: { type: :string }
            },
            required:   %w[content content_type]
          }
        }
      },
      required:   ['facts']
    },
    phase:   'gas_comprehend'
  )

  content = result.respond_to?(:message) ? result.message[:content] : result.to_s
  parsed = Legion::JSON.load(content)
  facts_array = parsed.is_a?(Hash) ? (parsed[:facts] || parsed['facts'] || []) : Array(parsed)
  facts_array.map do |f|
    {
      content:      f[:content] || f['content'],
      content_type: (f[:content_type] || f['content_type'] || 'fact').to_sym
    }
  end
rescue StandardError
  mechanical_comprehend(messages, response)
end

.llm_synthesize(facts) ⇒ Object 



204
205
206
207
208
209
210
211
212
213
214
215
216
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
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 204

def llm_synthesize(facts)
  facts_text = facts.each_with_index.map { |f, i| "[#{i}] (#{f[:content_type]}) #{f[:content]}" }.join("\n")

  prompt = <<~PROMPT
    Given these knowledge entries, generate derivative insights (inferences, implications, or connections).
    Each synthesis should combine information from multiple sources.

    Entries:
    #{facts_text}

    Return JSON with a "synthesis" array where each item has: content (string), content_type (inference/implication/connection), source_indices (array of entry indices used).
  PROMPT

  result = Legion::LLM::Pipeline::GaiaCaller.structured(
    message: prompt.strip,
    schema:  {
      type:       :object,
      properties: {
        synthesis: {
          type:  :array,
          items: {
            type:       :object,
            properties: {
              content:        { type: :string },
              content_type:   { type: :string },
              source_indices: { type: :array, items: { type: :integer } }
            },
            required:   %w[content content_type source_indices]
          }
        }
      },
      required:   ['synthesis']
    },
    phase:   'gas_synthesize'
  )

  content = result.respond_to?(:message) ? result.message[:content] : result.to_s
  parsed = Legion::JSON.load(content)
  items = parsed.is_a?(Hash) ? (parsed[:synthesis] || parsed['synthesis'] || []) : []

  items.map { |item| build_synthesis_entry(item, facts) }
rescue StandardError
  []
end

.mechanical_comprehend(_messages, response) ⇒ Object 



325
326
327
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 325

def mechanical_comprehend(_messages, response)
  [{ content: response, content_type: :observation }]
end

.phase_anticipate(facts, _synthesis) ⇒ Object

Phase 6: Anticipate - pre-cache likely follow-up questions



121
122
123
124
125
126
127
128
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 121

def phase_anticipate(facts, _synthesis)
  return [] if facts.empty?
  return [] unless llm_available?

  llm_anticipate(facts)
rescue StandardError
  []
end

.phase_comprehend(audit_event) ⇒ Object

Phase 1: Comprehend - extract typed facts from the exchange



39
40
41
42
43
44
45
46
47
48
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 39

def phase_comprehend(audit_event)
  messages = audit_event[:messages]
  response = audit_event[:response_content]

  if llm_available?
    llm_comprehend(messages, response)
  else
    mechanical_comprehend(messages, response)
  end
end

.phase_deposit(facts, _entities, _relations, _synthesis, audit_event) ⇒ Object

Phase 5: Deposit - atomic write to Apollo



97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 97

def phase_deposit(facts, _entities, _relations, _synthesis, audit_event)
  return { deposited: 0 } unless defined?(Runners::Knowledge)

  deposited = 0
  facts.each do |fact|
    Runners::Knowledge.handle_ingest(
      content:          fact[:content],
      content_type:     fact[:content_type].to_s,
      tags:             %w[gas auto_extracted],
      source_agent:     'gas_pipeline',
      source_provider:  audit_event.dig(:routing, :provider)&.to_s,
      knowledge_domain: 'general',
      context:          { source_request_id: audit_event[:request_id] }
    )
    deposited += 1
  rescue StandardError => e
    Legion::Logging.warn("GAS deposit error: #{e.message}") if defined?(Legion::Logging)
  end
  { deposited: deposited }
end

.phase_extract(audit_event, _facts) ⇒ Object

Phase 2: Extract - entity extraction (delegates to existing EntityExtractor)



51
52
53
54
55
56
57
58
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 51

def phase_extract(audit_event, _facts)
  return [] unless defined?(Runners::EntityExtractor)

  result = Runners::EntityExtractor.extract_entities(text: audit_event[:response_content])
  result[:success] ? (result[:entities] || []) : []
rescue StandardError
  []
end

.phase_relate(facts, _entities) ⇒ Object

Phase 3: Relate - classify relationships between new and existing entries



68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 68

def phase_relate(facts, _entities)
  return [] unless defined?(Runners::Knowledge)

  existing = fetch_similar_entries(facts)
  return [] if existing.empty?

  relations = []
  facts.each do |fact|
    existing.each do |entry|
      relation = classify_relation(fact, entry)
      relations << relation if relation
    end
  end
  relations
end

.phase_synthesize(facts, _relations) ⇒ Object

Phase 4: Synthesize - generate derivative knowledge



87
88
89
90
91
92
93
94
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 87

def phase_synthesize(facts, _relations)
  return [] if facts.length < 2
  return [] unless llm_available?

  llm_synthesize(facts)
rescue StandardError
  []
end

.process(audit_event) ⇒ Object 



10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 10

def process(audit_event)
  return { phases_completed: 0, reason: 'no content' } unless processable?(audit_event)

  facts = phase_comprehend(audit_event)
  entities = phase_extract(audit_event, facts)
  relations = phase_relate(facts, entities)
  synthesis = phase_synthesize(facts, relations)
  deposit_result = phase_deposit(facts, entities, relations, synthesis, audit_event)
  anticipations = phase_anticipate(facts, synthesis)

  {
    phases_completed: 6,
    facts:            facts.length,
    entities:         entities.length,
    relations:        relations.length,
    synthesis:        synthesis.length,
    deposited:        deposit_result,
    anticipations:    anticipations.length
  }
rescue StandardError => e
  Legion::Logging.warn("GAS pipeline error: #{e.message}") if defined?(Legion::Logging)
  { phases_completed: 0, error: e.message }
end

.processable?(event) ⇒ Boolean

Returns:

  • (Boolean) 


34
35
36
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 34

def processable?(event)
  event[:messages]&.any? == true && !event[:response_content].nil?
end

.promote_to_pattern_store(question:, facts:) ⇒ Object 



307
308
309
310
311
312
313
314
315
316
317
 
# File 'lib/legion/extensions/apollo/runners/gas.rb', line 307

def promote_to_pattern_store(question:, facts:)
  return unless defined?(Legion::Extensions::Agentic::TBI::PatternStore)

  Legion::Extensions::Agentic::TBI::PatternStore.promote_candidate(
    intent:     question,
    resolution: { source: 'gas_anticipate', facts: facts.map { |f| f[:content] } },
    confidence: 0.5
  )
rescue StandardError
  nil
end