Class: Raif::Llm

Inherits:
Object
  • Object
show all
Includes:
ActiveModel::Model, Concerns::Llms::MessageFormatting
Defined in:
app/models/raif/llm.rb

Constant Summary collapse

VALID_RESPONSE_FORMATS = 
[:text, :json, :html].freeze

Instance Attribute Summary collapse

Class Method Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(key:, api_name:, display_name: nil, model_provider_settings: {}, supported_provider_managed_tools: [], supports_native_tool_use: true, temperature: nil, max_completion_tokens: nil, input_token_cost: nil, output_token_cost: nil) ⇒ Llm

Returns a new instance of Llm.



26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
 
# File 'app/models/raif/llm.rb', line 26

def initialize(
  key:,
  api_name:,
  display_name: nil,
  model_provider_settings: {},
  supported_provider_managed_tools: [],
  supports_native_tool_use: true,
  temperature: nil,
  max_completion_tokens: nil,
  input_token_cost: nil,
  output_token_cost: nil
)
  @key = key
  @api_name = api_name
  @display_name = display_name
  @provider_settings = model_provider_settings
  @supports_native_tool_use = supports_native_tool_use
  @default_temperature = temperature || 0.7
  @default_max_completion_tokens = max_completion_tokens
  @input_token_cost = input_token_cost
  @output_token_cost = output_token_cost
  @supported_provider_managed_tools = supported_provider_managed_tools.map(&:to_s)
end

Instance Attribute Details

#api_name ⇒ Object

Returns the value of attribute api_name.



8
9
10
 
# File 'app/models/raif/llm.rb', line 8

def api_name
  @api_name
end

#default_max_completion_tokens ⇒ Object

Returns the value of attribute default_max_completion_tokens.



8
9
10
 
# File 'app/models/raif/llm.rb', line 8

def default_max_completion_tokens
  @default_max_completion_tokens
end

#default_temperature ⇒ Object

Returns the value of attribute default_temperature.



8
9
10
 
# File 'app/models/raif/llm.rb', line 8

def default_temperature
  @default_temperature
end

#display_name ⇒ Object

Returns the value of attribute display_name.



8
9
10
 
# File 'app/models/raif/llm.rb', line 8

def display_name
  @display_name
end

#input_token_cost ⇒ Object

Returns the value of attribute input_token_cost.



8
9
10
 
# File 'app/models/raif/llm.rb', line 8

def input_token_cost
  @input_token_cost
end

#key ⇒ Object

Returns the value of attribute key.



8
9
10
 
# File 'app/models/raif/llm.rb', line 8

def key
  @key
end

#output_token_cost ⇒ Object

Returns the value of attribute output_token_cost.



8
9
10
 
# File 'app/models/raif/llm.rb', line 8

def output_token_cost
  @output_token_cost
end

#provider_settings ⇒ Object

Returns the value of attribute provider_settings.



8
9
10
 
# File 'app/models/raif/llm.rb', line 8

def provider_settings
  @provider_settings
end

#supported_provider_managed_tools ⇒ Object

Returns the value of attribute supported_provider_managed_tools.



8
9
10
 
# File 'app/models/raif/llm.rb', line 8

def supported_provider_managed_tools
  @supported_provider_managed_tools
end

#supports_native_tool_use ⇒ Object Also known as: supports_native_tool_use?

Returns the value of attribute supports_native_tool_use.



8
9
10
 
# File 'app/models/raif/llm.rb', line 8

def supports_native_tool_use
  @supports_native_tool_use
end

Class Method Details

.batch_inference_cost_multiplier ⇒ Object

Multiplier applied to per-token costs when a model completion was resolved through this provider's Batch API. Defaults to 0.5 (50% discount), which is what both Anthropic and OpenAI charge for batch requests today.



258
259
260
 
# File 'app/models/raif/llm.rb', line 258

def self.batch_inference_cost_multiplier
  0.5
end

.cache_creation_input_token_cost_multiplier ⇒ Object

Multiplier applied to the base input_token_cost to derive the per-token cost for cache creation writes. Return nil when there is no write surcharge.



237
238
239
 
# File 'app/models/raif/llm.rb', line 237

def self.cache_creation_input_token_cost_multiplier
  nil
end

.cache_read_input_token_cost_multiplier ⇒ Object

Multiplier applied to the base input_token_cost to derive the per-token cost for cache reads. Return nil when the provider has no cache pricing.



231
232
233
 
# File 'app/models/raif/llm.rb', line 231

def self.cache_read_input_token_cost_multiplier
  nil
end

.prompt_tokens_include_cached_tokens? ⇒ Boolean

Override in subclasses to indicate whether prompt_tokens reported by the provider already include cached tokens as a subset (OpenAI, Google, OpenRouter) or whether cached tokens are reported separately and are additive to prompt_tokens (Anthropic, Bedrock).

Returns:

  • (Boolean) 


225
226
227
 
# File 'app/models/raif/llm.rb', line 225

def self.prompt_tokens_include_cached_tokens?
  true
end

.streaming_supported_for_key?(model_key) ⇒ Boolean

Whether streaming is supported for the given Raif model key. A model key is considered unsupported if it matches any entry in Raif.config.streaming_unsupported_model_keys (each entry may be a String, Symbol, or Regexp). Used by #chat to transparently fall back to the non-streaming path for models with known-broken streaming endpoints.

Returns:

  • (Boolean) 


59
60
61
62
63
64
65
66
67
68
 
# File 'app/models/raif/llm.rb', line 59

def self.streaming_supported_for_key?(model_key)
  entries = Array(Raif.config.streaming_unsupported_model_keys)
  key_str = model_key.to_s
  entries.none? do |entry|
    case entry
    when Regexp then entry.match?(key_str)
    else entry.to_s == key_str
    end
  end
end

.supports_batch_inference? ⇒ Boolean

Whether this provider supports submitting model completions via a Batch API. Override in subclasses by including Raif::Concerns::Llms::SupportsBatchInference, which sets this to true.

Returns:

  • (Boolean) 


244
245
246
 
# File 'app/models/raif/llm.rb', line 244

def self.supports_batch_inference?
  false
end

.valid_response_formats ⇒ Object 



217
218
219
 
# File 'app/models/raif/llm.rb', line 217

def self.valid_response_formats
  VALID_RESPONSE_FORMATS
end

Instance Method Details

#build_forced_tool_choice(tool_name) ⇒ Hash

Build the tool_choice parameter to force a specific tool to be called. Each provider implements this to return the correct format.

Parameters:

  • tool_name (String) —

    The name of the tool to force

Returns:

  • (Hash) —

    The tool_choice parameter for the provider's API

Raises:

  • (NotImplementedError) 


270
271
272
 
# File 'app/models/raif/llm.rb', line 270

def build_forced_tool_choice(tool_name)
  raise NotImplementedError, "#{self.class.name} must implement #build_forced_tool_choice"
end

#build_pending_model_completion(messages:, response_format: :text, available_model_tools: [], source: nil, system_prompt: nil, temperature: nil, max_completion_tokens: nil, tool_choice: nil, stream_response: false, allow_parallel_tool_calls: false, anthropic_prompt_caching_enabled: false, bedrock_prompt_caching_enabled: false, raif_model_completion_batch: nil, batch_custom_id: nil) ⇒ Raif::ModelCompletion

Builds and persists a Raif::ModelCompletion without performing the request. Used by #chat (which then calls perform_model_completion!) and by callers that want to defer execution -- e.g. submitting through a provider Batch API via Raif::Task.build_for_batch / Raif::Task#prepare_for_batch!.

Returns:



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
215
 
# File 'app/models/raif/llm.rb', line 188

def build_pending_model_completion(messages:, response_format: :text, available_model_tools: [], source: nil,
  system_prompt: nil, temperature: nil, max_completion_tokens: nil, tool_choice: nil,
  stream_response: false, allow_parallel_tool_calls: false, anthropic_prompt_caching_enabled: false,
  bedrock_prompt_caching_enabled: false, raif_model_completion_batch: nil, batch_custom_id: nil)
  temperature ||= default_temperature
  max_completion_tokens ||= default_max_completion_tokens

  model_completion = Raif::ModelCompletion.create!(
    messages: format_messages(messages),
    system_prompt: system_prompt,
    response_format: response_format,
    source: source,
    llm_model_key: key.to_s,
    model_api_name: api_name,
    temperature: temperature,
    max_completion_tokens: max_completion_tokens,
    available_model_tools: available_model_tools,
    tool_choice: tool_choice&.to_s,
    stream_response: stream_response,
    raif_model_completion_batch: raif_model_completion_batch,
    batch_custom_id: batch_custom_id
  )

  model_completion.allow_parallel_tool_calls = allow_parallel_tool_calls
  model_completion.anthropic_prompt_caching_enabled = anthropic_prompt_caching_enabled
  model_completion.bedrock_prompt_caching_enabled = bedrock_prompt_caching_enabled
  model_completion
end

#build_required_tool_choice ⇒ Hash, String

Build the tool_choice parameter to require the model to call any tool (but not a specific one). Each provider implements this to return the correct format.

Returns:

  • (Hash, String) —

    The tool_choice parameter for the provider's API

Raises:

  • (NotImplementedError) 


277
278
279
 
# File 'app/models/raif/llm.rb', line 277

def build_required_tool_choice
  raise NotImplementedError, "#{self.class.name} must implement #build_required_tool_choice"
end

#chat(message: nil, messages: nil, response_format: :text, available_model_tools: [], source: nil, system_prompt: nil, temperature: nil, max_completion_tokens: nil, tool_choice: nil, allow_parallel_tool_calls: false, anthropic_prompt_caching_enabled: false, bedrock_prompt_caching_enabled: false, &block) ⇒ Object 



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
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
167
168
169
170
171
172
173
174
175
176
 
# File 'app/models/raif/llm.rb', line 74

def chat(message: nil, messages: nil, response_format: :text, available_model_tools: [], source: nil, system_prompt: nil, temperature: nil,
  max_completion_tokens: nil, tool_choice: nil, allow_parallel_tool_calls: false, anthropic_prompt_caching_enabled: false,
  bedrock_prompt_caching_enabled: false, &block)
  unless response_format.is_a?(Symbol)
    raise ArgumentError,
      "Raif::Llm#chat - Invalid response format: #{response_format}. Must be a symbol (you passed #{response_format.class}) and be one of: #{VALID_RESPONSE_FORMATS.join(", ")}" # rubocop:disable Layout/LineLength
  end

  unless VALID_RESPONSE_FORMATS.include?(response_format)
    raise ArgumentError, "Raif::Llm#chat - Invalid response format: #{response_format}. Must be one of: #{VALID_RESPONSE_FORMATS.join(", ")}"
  end

  unless message.present? || messages.present?
    raise ArgumentError, "Raif::Llm#chat - You must provide either a message: or messages: argument"
  end

  if message.present? && messages.present?
    raise ArgumentError, "Raif::Llm#chat - You must provide either a message: or messages: argument, not both"
  end

  # Normalize :required / "required" to the symbol form for validation
  tool_choice = :required if tool_choice.to_s == "required"

  if tool_choice == :required
    if available_model_tools.blank?
      raise ArgumentError,
        "Raif::Llm#chat - tool_choice: :required requires at least one available model tool"
    end
  elsif tool_choice.present? && !available_model_tools.map(&:to_s).include?(tool_choice.to_s)
    raise ArgumentError,
      "Raif::Llm#chat - Invalid tool choice: #{tool_choice} is not included in the available model tools: #{available_model_tools.join(", ")}"
  end

  # Runs before the ModelCompletion is created or any provider call is made,
  # and before the llm_api_requests_enabled guard so authorization applies
  # even when API requests are disabled. Vetoes by raising. Any raised
  # exception is tagged with Raif::Errors::ModelCompletionAuthorizationError
  # so wrapped flows (Raif::Task.run, Raif::Conversation) re-raise it to the
  # caller instead of swallowing it as an ordinary model failure.
  if Raif.config.model_completion_authorizer
    begin
      Raif.config.model_completion_authorizer.call(llm: self, source: source)
    rescue StandardError => e
      e.extend(Raif::Errors::ModelCompletionAuthorizationError) unless e.is_a?(Raif::Errors::ModelCompletionAuthorizationError)
      raise
    end
  end

  unless Raif.config.llm_api_requests_enabled
    Raif.logger.warn("LLM API requests are disabled. Skipping request to #{api_name}.")
    return
  end

  messages = [{ "role" => "user", "content" => message }] if message.present?

  temperature ||= default_temperature
  max_completion_tokens ||= default_max_completion_tokens

  stream_response = block_given? && streaming_supported?
  if block_given? && !stream_response
    Raif.logger.info(
      "Raif::Llm#chat: streaming requested but disabled for model key #{key.inspect} " \
        "via Raif.config.streaming_unsupported_model_keys; falling back to non-streaming."
    )
  end

  model_completion = build_pending_model_completion(
    messages: messages,
    response_format: response_format,
    available_model_tools: available_model_tools,
    source: source,
    system_prompt: system_prompt,
    temperature: temperature,
    max_completion_tokens: max_completion_tokens,
    tool_choice: tool_choice,
    stream_response: stream_response,
    allow_parallel_tool_calls: allow_parallel_tool_calls,
    anthropic_prompt_caching_enabled: anthropic_prompt_caching_enabled,
    bedrock_prompt_caching_enabled: bedrock_prompt_caching_enabled
  )

  model_completion.started!

  retry_with_backoff(model_completion) do
    perform_model_completion!(model_completion, &block)
    ensure_model_completion_present!(model_completion)
  end

  model_completion.completed!
  model_completion
rescue Raif::Errors::StreamingError => e
  Rails.logger.error("Raif streaming error -- code: #{e.code} -- type: #{e.type} -- message: #{e.message} -- event: #{e.event}")
  model_completion&.record_failure!(e) unless model_completion&.failed?
  raise e
rescue Faraday::Error => e
  Raif.logger.error("LLM API request failed (status: #{e.response_status}): #{e.message}")
  Raif.logger.error(e.response_body)
  model_completion&.record_failure!(e) unless model_completion&.failed?
  raise e
rescue StandardError => e
  model_completion&.record_failure!(e) unless model_completion&.failed?
  raise e
end

#name ⇒ Object 



50
51
52
 
# File 'app/models/raif/llm.rb', line 50

def name
  I18n.t("raif.model_names.#{key}", default: display_name || key.to_s.humanize)
end

#perform_model_completion!(model_completion, &block) ⇒ Object

Raises:

  • (NotImplementedError) 


178
179
180
 
# File 'app/models/raif/llm.rb', line 178

def perform_model_completion!(model_completion, &block)
  raise NotImplementedError, "#{self.class.name} must implement #perform_model_completion!"
end

#streaming_supported? ⇒ Boolean

Returns:

  • (Boolean) 


70
71
72
 
# File 'app/models/raif/llm.rb', line 70

def streaming_supported?
  self.class.streaming_supported_for_key?(key)
end

#supports_batch_inference? ⇒ Boolean

Instance-level shortcut for the class-level predicate so callers can use the idiomatic Raif.llm(:some_key).supports_batch_inference? form instead of reaching through to the class.

Returns:

  • (Boolean) 


251
252
253
 
# File 'app/models/raif/llm.rb', line 251

def supports_batch_inference?
  self.class.supports_batch_inference?
end

#supports_faithful_required_tool_choice?(available_model_tools) ⇒ Boolean

Whether the provider can faithfully enforce tool_choice: :required for the given tool set. Override in subclasses when a provider can only enforce required tool use for some tool types.

Returns:

  • (Boolean) 


284
285
286
 
# File 'app/models/raif/llm.rb', line 284

def supports_faithful_required_tool_choice?(available_model_tools)
  available_model_tools.present?
end

#supports_parallel_tool_calls? ⇒ Boolean

Whether this model can handle being asked to make multiple tool calls in a single response. Override (per provider or per model key) to return false for models that reject the parallel-tool-call request parameter or that produce worse results when allowed to batch. Agents consult this before enabling parallel tool calls; when false they fall back to one tool call per step.

Returns:

  • (Boolean) 


293
294
295
 
# File 'app/models/raif/llm.rb', line 293

def supports_parallel_tool_calls?
  true
end

#supports_provider_managed_tool?(tool_klass) ⇒ Boolean

Returns:

  • (Boolean) 


262
263
264
 
# File 'app/models/raif/llm.rb', line 262

def supports_provider_managed_tool?(tool_klass)
  supported_provider_managed_tools&.include?(tool_klass.to_s)
end

#validate_provider_managed_tool_support!(tool) ⇒ Object 



297
298
299
300
301
302
 
# File 'app/models/raif/llm.rb', line 297

def validate_provider_managed_tool_support!(tool)
  unless supports_provider_managed_tool?(tool)
    raise Raif::Errors::UnsupportedFeatureError,
      "Invalid provider-managed tool: #{tool.name} for #{key}"
  end
end