Module: Legion::LLM::Call::Embeddings

Extended by:
Legion::Logging::Helper
Defined in:
lib/legion/llm/call/embeddings.rb

Constant Summary collapse

PREFIX_REGISTRY = 
{
  'nomic-embed-text'  => { document: 'search_document: ', query: 'search_query: ' },
  'mxbai-embed-large' => { query: 'Represent this sentence for searching relevant passages: ' }
}.freeze

Class Method Summary collapse

Class Method Details

.default_model ⇒ Object

G15: returns the configured embedding model (pinned). nil if not configured. Reads :default_model first, falls back to the deprecated :model alias.



170
171
172
 
# File 'lib/legion/llm/call/embeddings.rb', line 170

def default_model
  configured_default_model
end

.generate(text:, model: nil, **opts) ⇒ Object

G15: Embedding callers go through Router.request_lane(type: :embedding, …). Strict pin on (provider, instance, model) when configured — no cross-model failover (vector-comparability preserved). A down pinned lane → NoLaneAvailable (400, not silent dimension-switch).



23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
 
# File 'lib/legion/llm/call/embeddings.rb', line 23

def generate(text:, model: nil, **opts)
  return not_started_result(model, nil) unless LLM.started?

  pinned_model    = model || configured_default_model
  pinned_provider = configured_provider
  pinned_instance = configured_instance
  if pinned_model.nil? || pinned_model.to_s.empty?
    raise Legion::LLM::Errors::ConfigError,
          'no embedding model configured — set :llm, :embedding, :default_model in settings'
  end

  lane = Legion::LLM::Router.request_lane(
    type:      :embedding,
    models:    [pinned_model.to_s],
    providers: pinned_provider ? [pinned_provider.to_sym] : [],
    instances: pinned_instance ? [pinned_instance.to_sym] : []
  )
  if lane.nil?
    raise Legion::LLM::Errors::NoLaneAvailable.new(
      filters: { type: :embedding, models: [pinned_model],
                 providers: pinned_provider ? [pinned_provider] : [],
                 instances: pinned_instance ? [pinned_instance] : [] }
    )
  end

  provider = lane[:provider_family]
  instance = lane[:instance_id]
  lane_model = lane[:model]

  text = coerce_text(text)
  dimensions = opts[:dimensions]
  task       = opts[:task] || :document
  prepared_texts = prepare_embedding_texts(text, provider: provider, model: lane_model, task: task)
  dispatch_text  = prepared_texts.one? ? prepared_texts.first : prepared_texts

  log.info("[llm][embed] action=generate provider=#{provider} instance=#{instance || 'default'} " \
           "model=#{lane_model} task=#{task} text_chars=#{text.length} chunks=#{prepared_texts.size}")

  started_at = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
  response = Dispatch.call(
    provider:   provider,
    instance:   instance,
    capability: :embed,
    model:      lane_model,
    text:       dispatch_text,
    dimensions: dimensions
  )
  elapsed = ((::Process.clock_gettime(::Process::CLOCK_MONOTONIC) - started_at) * 1000).round(1)

  vector = if prepared_texts.size > 1
             aggregate_vectors(response[:result],
                               weights:  prepared_texts.map(&:length),
                               model:    lane_model,
                               provider: provider)
           else
             normalize_vector(response[:result])
           end
  vector = enforce_dimensions(vector) if enforce_dimension?
  tokens = extract_tokens(response)

  log.info("[llm][embed] action=generate.complete provider=#{provider} " \
           "instance=#{instance || 'default'} model=#{lane_model} " \
           "dimensions=#{vector&.size || 0} tokens=#{tokens} chunks=#{prepared_texts.size} duration_ms=#{elapsed}")

  {
    vector:     vector,
    model:      lane_model,
    provider:   provider,
    dimensions: vector&.size || 0,
    tokens:     tokens,
    chunks:     prepared_texts.size
  }
rescue Legion::LLM::Errors::NoLaneAvailable, Legion::LLM::Errors::ConfigError, Legion::LLM::LLMError
  raise
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.embeddings.generate')
  { vector: nil, model: pinned_model, provider: nil, error: e.message }
end

.generate_batch(texts:, model: nil, dimensions: nil, task: :document) ⇒ Object 



102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
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
 
# File 'lib/legion/llm/call/embeddings.rb', line 102

def generate_batch(texts:, model: nil, dimensions: nil, task: :document, **)
  return texts.map { { vector: nil, error: 'LLM not started' } } unless LLM.started?

  pinned_model    = model || configured_default_model
  pinned_provider = configured_provider
  pinned_instance = configured_instance
  return texts.map { { vector: nil, error: 'no embedding model configured' } } if pinned_model.nil? || pinned_model.to_s.empty?

  lane = Legion::LLM::Router.request_lane(
    type:      :embedding,
    models:    [pinned_model.to_s],
    providers: pinned_provider ? [pinned_provider.to_sym] : [],
    instances: pinned_instance ? [pinned_instance.to_sym] : []
  )
  if lane.nil?
    raise Legion::LLM::Errors::NoLaneAvailable.new(
      filters: { type: :embedding, models: [pinned_model],
                 providers: pinned_provider ? [pinned_provider] : [],
                 instances: pinned_instance ? [pinned_instance] : [] }
    )
  end

  provider = lane[:provider_family]
  instance = lane[:instance_id]
  model    = lane[:model]

  log.info("[llm][embed] action=generate_batch provider=#{provider} instance=#{instance || 'default'} " \
           "model=#{model} count=#{texts.size} task=#{task}")

  raw_texts = texts.map { |t| coerce_text(t) }
  prepared_texts = raw_texts.map { |t| prepare_embedding_texts(t, provider: provider, model: model, task: task) }
  if prepared_texts.any? { |chunks| chunks.size > 1 }
    return generate_chunked_batch(
      raw_texts,
      model:      model,
      provider:   provider,
      instance:   instance,
      dimensions: dimensions,
      task:       task
    )
  end

  texts = prepared_texts.map(&:first)

  started_at = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
  response = Dispatch.call(
    provider:   provider,
    instance:   instance,
    capability: :embed,
    model:      model,
    text:       texts,
    dimensions: dimensions
  )
  elapsed = ((::Process.clock_gettime(::Process::CLOCK_MONOTONIC) - started_at) * 1000).round(1)

  result = normalize_batch(response[:result], model, provider)

  log.info("[llm][embed] action=generate_batch.complete provider=#{provider} " \
           "model=#{model} count=#{result.size} duration_ms=#{elapsed}")

  result
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.embeddings.generate_batch')
  texts.map { { vector: nil, model: model, provider: provider, error: e.message } }
end