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 



117
118
119
 
# File 'lib/legion/llm/call/embeddings.rb', line 117

def default_model
  resolve_model
end

.generate(text:, model: nil, provider: nil, instance: nil, dimensions: nil, task: :document) ⇒ Object 



19
20
21
22
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
 
# File 'lib/legion/llm/call/embeddings.rb', line 19

def generate(text:, model: nil, provider: nil, instance: nil,
             dimensions: nil, task: :document)
  return not_started_result(model, provider) unless LLM.started?

  provider ||= resolve_provider
  return unavailable_result(model, provider) unless provider

  model ||= resolve_model
  text = coerce_text(text)
  text_length = text.length
  prepared_texts = prepare_embedding_texts(text, provider: provider, model: 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=#{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:      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: 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=#{model} dimensions=#{vector&.size || 0} tokens=#{tokens} chunks=#{prepared_texts.size} duration_ms=#{elapsed}")

  {
    vector:     vector,
    model:      model,
    provider:   provider,
    dimensions: vector&.size || 0,
    tokens:     tokens,
    chunks:     prepared_texts.size
  }
rescue StandardError => e
  handle_exception(e, level: :warn, operation: 'llm.embeddings.generate')
  { vector: nil, model: model, provider: provider, error: e.message }
end

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



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
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
 
# File 'lib/legion/llm/call/embeddings.rb', line 70

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

  provider ||= resolve_provider
  model ||= resolve_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