Class: DSPy::LM::Adapters::OpenAI::SchemaConverter

Inherits:
Object
  • Object
show all
Extended by:
T::Sig
Defined in:
lib/dspy/lm/adapters/openai/schema_converter.rb

Overview

Converts DSPy signatures to OpenAI structured output format

Constant Summary collapse

STRUCTURED_OUTPUT_MODELS =

Models that support structured outputs as of July 2025

T.let([
  "gpt-4o-mini",
  "gpt-4o-2024-08-06",
  "gpt-4o",
  "gpt-4-turbo",
  "gpt-4-turbo-2024-04-09"
].freeze, T::Array[String])

Class Method Summary collapse

Class Method Details

.all_have_discriminators?(schemas) ⇒ Boolean

Returns:

  • (Boolean) 


111
112
113
114
115
116
117
118
119
 
# File 'lib/dspy/lm/adapters/openai/schema_converter.rb', line 111

def self.all_have_discriminators?(schemas)
  schemas.all? do |schema|
    next false unless schema.is_a?(Hash)
    next false unless schema[:properties].is_a?(Hash)
    
    # Check if any property has a const value (our discriminator pattern)
    schema[:properties].any? { |_, prop| prop.is_a?(Hash) && prop[:const] }
  end
end

.convert_oneof_to_anyof_if_safe(schema) ⇒ Object 



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
101
102
103
104
105
106
107
 
# File 'lib/dspy/lm/adapters/openai/schema_converter.rb', line 63

def self.convert_oneof_to_anyof_if_safe(schema)
  return schema unless schema.is_a?(Hash)
  
  result = schema.dup
  
  # Check if this schema has oneOf that we can safely convert
  if result[:oneOf]
    if all_have_discriminators?(result[:oneOf])
      # Safe to convert - discriminators ensure mutual exclusivity
      result[:anyOf] = result.delete(:oneOf).map { |s| convert_oneof_to_anyof_if_safe(s) }
    else
      # Unsafe conversion - raise error
      raise DSPy::UnsupportedSchemaError.new(
        "OpenAI structured outputs do not support oneOf schemas without discriminator fields. " \
        "The schema contains union types that cannot be safely converted to anyOf. " \
        "Please use enhanced_prompting strategy instead or add discriminator fields to union types."
      )
    end
  end
  
  # Recursively process nested schemas
  if result[:properties].is_a?(Hash)
    result[:properties] = result[:properties].transform_values { |v| convert_oneof_to_anyof_if_safe(v) }
  end
  
  if result[:items].is_a?(Hash)
    result[:items] = convert_oneof_to_anyof_if_safe(result[:items])
  end
  
  # Process arrays of schema items
  if result[:items].is_a?(Array)
    result[:items] = result[:items].map { |item| 
      item.is_a?(Hash) ? convert_oneof_to_anyof_if_safe(item) : item 
    }
  end
  
  # Process anyOf arrays (in case there are nested oneOf within anyOf)
  if result[:anyOf].is_a?(Array)
    result[:anyOf] = result[:anyOf].map { |item| 
      item.is_a?(Hash) ? convert_oneof_to_anyof_if_safe(item) : item 
    }
  end
  
  result
end

.supports_structured_outputs?(model) ⇒ Boolean

Returns:

  • (Boolean) 


122
123
124
125
126
127
128
 
# File 'lib/dspy/lm/adapters/openai/schema_converter.rb', line 122

def self.supports_structured_outputs?(model)
  # Extract base model name without provider prefix
  base_model = model.sub(/^openai\//, "")
  
  # Check if it's a supported model or a newer version
  STRUCTURED_OUTPUT_MODELS.any? { |supported| base_model.start_with?(supported) }
end

.to_openai_format(signature_class, name: nil, strict: true) ⇒ Object 



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
 
# File 'lib/dspy/lm/adapters/openai/schema_converter.rb', line 23

def self.to_openai_format(signature_class, name: nil, strict: true)
  # Get the output JSON schema from the signature class
  output_schema = signature_class.output_json_schema
  
  # Convert oneOf to anyOf where safe, or raise error for unsupported cases
  output_schema = convert_oneof_to_anyof_if_safe(output_schema)
  
  # Build the complete schema with OpenAI-specific modifications
  dspy_schema = {
    "$schema": "http://json-schema.org/draft-06/schema#",
    type: "object",
    properties: output_schema[:properties] || {},
    required: openai_required_fields(signature_class, output_schema)
  }

  # Generate a schema name if not provided
  schema_name = name || generate_schema_name(signature_class)

  # Remove the $schema field as OpenAI doesn't use it
  openai_schema = dspy_schema.except(:$schema)

  # Add additionalProperties: false for strict mode and fix nested struct schemas
  if strict
    openai_schema = add_additional_properties_recursively(openai_schema)
    openai_schema = fix_nested_struct_required_fields(openai_schema)
  end

  # Wrap in OpenAI's required format
  {
    type: "json_schema",
    json_schema: {
      name: schema_name,
      strict: strict,
      schema: openai_schema
    }
  }
end

.validate_compatibility(schema) ⇒ Object 



131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
 
# File 'lib/dspy/lm/adapters/openai/schema_converter.rb', line 131

def self.validate_compatibility(schema)
  issues = []

  # Check for deeply nested objects (OpenAI has depth limits)
  depth = calculate_depth(schema)
  if depth > 5
    issues << "Schema depth (#{depth}) exceeds recommended limit of 5 levels"
  end

  # Check for unsupported JSON Schema features
  if contains_pattern_properties?(schema)
    issues << "Pattern properties are not supported in OpenAI structured outputs"
  end

  if contains_conditional_schemas?(schema)
    issues << "Conditional schemas (if/then/else) are not supported"
  end

  issues
end