2054db016c
* Add support for O3-mini - Add support for the o3 mini set of models, including tests that match the constraints in OpenAI's API docs (https://platform.openai.com/docs/models#o3-mini). * Deprecate and refactor - Deprecate `ErrO1BetaLimitationsLogprobs` and `ErrO1BetaLimitationsOther` - Implement `validationRequestForReasoningModels`, which works on both o1 & o3, and has per-model-type restrictions on functionality (eg, o3 class are allowed function calls and system messages, o1 isn't) * Move reasoning validation to `reasoning_validator.go` - Add a `NewReasoningValidator` which exposes a `Validate()` method for a given request - Also adds a test for chat streams * Final nits
112 lines
4.2 KiB
Go
112 lines
4.2 KiB
Go
package openai
|
|
|
|
import (
|
|
"errors"
|
|
"strings"
|
|
)
|
|
|
|
var (
|
|
// Deprecated: use ErrReasoningModelMaxTokensDeprecated instead.
|
|
ErrO1MaxTokensDeprecated = errors.New("this model is not supported MaxTokens, please use MaxCompletionTokens") //nolint:lll
|
|
ErrCompletionUnsupportedModel = errors.New("this model is not supported with this method, please use CreateChatCompletion client method instead") //nolint:lll
|
|
ErrCompletionStreamNotSupported = errors.New("streaming is not supported with this method, please use CreateCompletionStream") //nolint:lll
|
|
ErrCompletionRequestPromptTypeNotSupported = errors.New("the type of CompletionRequest.Prompt only supports string and []string") //nolint:lll
|
|
)
|
|
|
|
var (
|
|
ErrO1BetaLimitationsMessageTypes = errors.New("this model has beta-limitations, user and assistant messages only, system messages are not supported") //nolint:lll
|
|
ErrO1BetaLimitationsTools = errors.New("this model has beta-limitations, tools, function calling, and response format parameters are not supported") //nolint:lll
|
|
// Deprecated: use ErrReasoningModelLimitations* instead.
|
|
ErrO1BetaLimitationsLogprobs = errors.New("this model has beta-limitations, logprobs not supported") //nolint:lll
|
|
ErrO1BetaLimitationsOther = errors.New("this model has beta-limitations, temperature, top_p and n are fixed at 1, while presence_penalty and frequency_penalty are fixed at 0") //nolint:lll
|
|
)
|
|
|
|
var (
|
|
//nolint:lll
|
|
ErrReasoningModelMaxTokensDeprecated = errors.New("this model is not supported MaxTokens, please use MaxCompletionTokens")
|
|
ErrReasoningModelLimitationsLogprobs = errors.New("this model has beta-limitations, logprobs not supported") //nolint:lll
|
|
ErrReasoningModelLimitationsOther = errors.New("this model has beta-limitations, temperature, top_p and n are fixed at 1, while presence_penalty and frequency_penalty are fixed at 0") //nolint:lll
|
|
)
|
|
|
|
var unsupportedToolsForO1Models = map[ToolType]struct{}{
|
|
ToolTypeFunction: {},
|
|
}
|
|
|
|
var availableMessageRoleForO1Models = map[string]struct{}{
|
|
ChatMessageRoleUser: {},
|
|
ChatMessageRoleAssistant: {},
|
|
}
|
|
|
|
// ReasoningValidator handles validation for o-series model requests.
|
|
type ReasoningValidator struct{}
|
|
|
|
// NewReasoningValidator creates a new validator for o-series models.
|
|
func NewReasoningValidator() *ReasoningValidator {
|
|
return &ReasoningValidator{}
|
|
}
|
|
|
|
// Validate performs all validation checks for o-series models.
|
|
func (v *ReasoningValidator) Validate(request ChatCompletionRequest) error {
|
|
o1Series := strings.HasPrefix(request.Model, "o1")
|
|
o3Series := strings.HasPrefix(request.Model, "o3")
|
|
|
|
if !o1Series && !o3Series {
|
|
return nil
|
|
}
|
|
|
|
if err := v.validateReasoningModelParams(request); err != nil {
|
|
return err
|
|
}
|
|
|
|
if o1Series {
|
|
if err := v.validateO1Specific(request); err != nil {
|
|
return err
|
|
}
|
|
}
|
|
|
|
return nil
|
|
}
|
|
|
|
// validateReasoningModelParams checks reasoning model parameters.
|
|
func (v *ReasoningValidator) validateReasoningModelParams(request ChatCompletionRequest) error {
|
|
if request.MaxTokens > 0 {
|
|
return ErrReasoningModelMaxTokensDeprecated
|
|
}
|
|
if request.LogProbs {
|
|
return ErrReasoningModelLimitationsLogprobs
|
|
}
|
|
if request.Temperature > 0 && request.Temperature != 1 {
|
|
return ErrReasoningModelLimitationsOther
|
|
}
|
|
if request.TopP > 0 && request.TopP != 1 {
|
|
return ErrReasoningModelLimitationsOther
|
|
}
|
|
if request.N > 0 && request.N != 1 {
|
|
return ErrReasoningModelLimitationsOther
|
|
}
|
|
if request.PresencePenalty > 0 {
|
|
return ErrReasoningModelLimitationsOther
|
|
}
|
|
if request.FrequencyPenalty > 0 {
|
|
return ErrReasoningModelLimitationsOther
|
|
}
|
|
|
|
return nil
|
|
}
|
|
|
|
// validateO1Specific checks O1-specific limitations.
|
|
func (v *ReasoningValidator) validateO1Specific(request ChatCompletionRequest) error {
|
|
for _, m := range request.Messages {
|
|
if _, found := availableMessageRoleForO1Models[m.Role]; !found {
|
|
return ErrO1BetaLimitationsMessageTypes
|
|
}
|
|
}
|
|
|
|
for _, t := range request.Tools {
|
|
if _, found := unsupportedToolsForO1Models[t.Type]; found {
|
|
return ErrO1BetaLimitationsTools
|
|
}
|
|
}
|
|
return nil
|
|
}
|