maritaca

Documentation

Index

• Variables
• type LLM
  • func New(opts ...Option) (*LLM, error)
  • func (o *LLM) Call(ctx context.Context, prompt string, options ...llms.CallOption) (string, error)
  • func (o *LLM) GenerateContent(ctx context.Context, messages []llms.MessageContent, ...) (*llms.ContentResponse, error)
• type Option
  • func WithChatMode(chatMode bool) Option
  • func WithCustomTemplate(template string) Option
  • func WithDoSample(doSample bool) Option
  • func WithFormat(format string) Option
  • func WithHTTPClient(client *http.Client) Option
  • func WithLangsmithParentId(parentId string) Option
  • func WithLangsmithgoRunId(runId string) Option
  • func WithMaxTokens(maxTokens int) Option
  • func WithModel(model string) Option
  • func WithRepetitionPenalty(repetitionPenalty float64) Option
  • func WithServerURL(rawURL string) Option
  • func WithStoppingTokens(tokens []string) Option
  • func WithStream(stream bool) Option
  • func WithSystemPrompt(p string) Option
  • func WithTemperature(temperature float64) Option
  • func WithToken(token string) Option
  • func WithTokensPerMessage(tokensPerMessage int) Option
  • func WithTopP(topP float64) Option

Variables

var (
	ErrEmptyResponse       = errors.New("no response")
	ErrIncompleteEmbedding = errors.New("no all input got emmbedded")
)

Types

type LLM

type LLM struct {
	CallbacksHandler callbacks.Handler
	// contains filtered or unexported fields
}

LLM is a maritaca LLM implementation.

func New

func New(opts ...Option) (*LLM, error)

New creates a new maritaca LLM implementation.

func (*LLM) Call

func (o *LLM) Call(ctx context.Context, prompt string, options ...llms.CallOption) (string, error)

Call Implement the call interface for LLM.

func (*LLM) GenerateContent

func (o *LLM) GenerateContent(ctx context.Context, messages []llms.MessageContent, options ...llms.CallOption) (*llms.ContentResponse, error)

GenerateContent implements the Model interface. nolint: goerr113

type Option

type Option func(*options)

func WithChatMode

func WithChatMode(chatMode bool) Option

WithChatMode Set the chat mode. default: true If True, the model will run in chat mode, where messages is a string containing the user's message or a list of messages containing the iterations of the conversation between user and assistant. If False, messages must be a string containing the desired prompt.

func WithCustomTemplate

func WithCustomTemplate(template string) Option

WithCustomTemplate To override the templating done on maritaca model side.

func WithDoSample

func WithDoSample(doSample bool) Option

WithDoSample Set the model's generation will be sampled via top-k sampling. Default: true If True, the model's generation will be sampled via top-k sampling. Otherwise, the generation will always select the token with the highest probability. Using do_sample=False leads to a deterministic result, but with less diversity.

func WithFormat

func WithFormat(format string) Option

WithFormat Sets the maritaca output format (currently maritaca only supports "json").

func WithHTTPClient

func WithHTTPClient(client *http.Client) Option

WithHTTPClient Set custom http client.

func WithLangsmithParentId

func WithLangsmithParentId(parentId string) Option

WithLangsmithParentId Set the langsmith parent id.

func WithLangsmithgoRunId

func WithLangsmithgoRunId(runId string) Option

WithLangsmithgoRunId Set the langsmithgo run id.

func WithMaxTokens

func WithMaxTokens(maxTokens int) Option

WithMaxTokens Set the maximum number of tokens that will be generated by the mode. minimum: 1 Maximum number of tokens that will be generated by the mode.

func WithModel

func WithModel(model string) Option

WithModel Set the model to use.

func WithRepetitionPenalty

func WithRepetitionPenalty(repetitionPenalty float64) Option

WithFrequencyPenalty Set the frequency penalty.

minimum: 0

default: 1 Repetition penalty. Positive values encourage the model not to repeat previously generated tokens.

func WithServerURL

func WithServerURL(rawURL string) Option

WithServerURL Set the URL of the maritaca instance to use.

func WithStoppingTokens

func WithStoppingTokens(tokens []string) Option

WithFrequencyPenalty Set the frequency penalty. List of tokens that, when generated, indicate that the model should stop generating tokens.

func WithStream

func WithStream(stream bool) Option

WithStream Set the model will run in streaming mode. default: false If True, the model will run in streaming mode, where tokens will be generated and returned to the client as they are produced. If False, the model will run in batch mode, where all tokens will be generated before being returned to the client.

func WithSystemPrompt

func WithSystemPrompt(p string) Option

WithSystem Set the system prompt. This is only valid if WithCustomTemplate is not set and the maritaca model use .System in its model template OR if WithCustomTemplate is set using {{.System}}.

func WithTemperature

func WithTemperature(temperature float64) Option

WithTemperature Set the sampling temperature. minimum: 0 default: 0.7 Sampling temperature (greater than or equal to zero). Higher values lead to greater diversity in generation but also increase the likelihood of generating nonsensical texts. Values closer to zero result in more plausible texts but increase the chances of generating repetitive texts.

func WithToken

func WithToken(token string) Option

WithToken Set the token to use. token use as key.

func WithTokensPerMessage

func WithTokensPerMessage(tokensPerMessage int) Option

WithTokensPerMessage Set the number of tokens that will be returned per message.

minimum: 1

default: 4 Number of tokens that will be returned per message. This field is ignored if stream=False.

func WithTopP

func WithTopP(topP float64) Option

WithTopK Set the number of top tokens to consider for sampling. exclusiveMaximum: 1 exclusiveMinimum: 0 default: 0.95 If less than 1, it retains only the top tokens with cumulative probability >= top_p (nucleus filtering). For example, 0.95 means that only the tokens that make up the top 95% of the probability mass are considered when predicting the next token.

Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751).