Merge pull request #494 from innogames/md_openai_config

openai: add more config flags

Author: brainexe

command/openai/api.go

@@ -13,21 +13,20 @@ const (
 	roleAssistant = "assistant"
 )
 
-// https://platform.openai.com/docs/guides/chat/chat-completions-beta
 // https://platform.openai.com/docs/api-reference/chat
 type ChatRequest struct {
-	Model            string         `json:"model"`
-	Messages         []ChatMessage  `json:"messages"`
-	Temperature      float32        `json:"temperature,omitempty"`
-	TopP             float32        `json:"top_p,omitempty"`
-	N                int            `json:"n,omitempty"`
-	Stop             []string       `json:"stop,omitempty"`
-	Stream           bool           `json:"stream,omitempty"`
-	MaxTokens        int            `json:"max_tokens,omitempty"`
-	PresencePenalty  float32        `json:"presence_penalty,omitempty"`
-	FrequencyPenalty float32        `json:"frequency_penalty,omitempty"`
-	LogitBias        map[string]int `json:"logit_bias,omitempty"`
-	User             string         `json:"user,omitempty"`
+	Model            string        `json:"model"`
+	Messages         []ChatMessage `json:"messages"`
+	Temperature      float32       `json:"temperature,omitempty"`
+	TopP             float32       `json:"top_p,omitempty"`
+	N                int           `json:"n,omitempty"`
+	Stop             []string      `json:"stop,omitempty"`
+	Stream           bool          `json:"stream,omitempty"`
+	MaxTokens        int           `json:"max_tokens,omitempty"`
+	PresencePenalty  float32       `json:"presence_penalty,omitempty"`
+	FrequencyPenalty float32       `json:"frequency_penalty,omitempty"`
+	User             string        `json:"user,omitempty"`
+	Seed             string        `json:"seed,omitempty"`
 }
 
 type ChatMessage struct {

command/openai/client.go

@@ -18,6 +18,8 @@ func CallChatGPT(cfg Config, inputMessages []ChatMessage, stream bool) (<-chan s
 	jsonData, _ := json.Marshal(ChatRequest{
 		Model:       cfg.Model,
 		Temperature: cfg.Temperature,
+		Seed:        cfg.Seed,
+		MaxTokens:   cfg.MaxTokens,
 		Stream:      stream,
 		Messages:    inputMessages,
 	})

command/openai/command.go

@@ -101,7 +101,9 @@ func (c *chatGPTCommand) startConversation(message msg.Ref, text string) bool {
 			})
 		}
 		storageIdentifier = getIdentifier(message.GetChannel(), message.GetThread())
-		log.Infof("openai thread context: %s", messageHistory)
+		if c.cfg.LogTexts {
+			log.Infof("openai thread context: %s", messageHistory)
+		}
 	} else if linkRe.MatchString(text) {
 		// a link to another thread was posted -> use this messages as context
 		link := linkRe.FindStringSubmatch(text)
@@ -242,14 +244,19 @@ func (c *chatGPTCommand) callAndStore(messages []ChatMessage, storageIdentifier
 		stats.Increase("openai_output_tokens", estimateTokensForMessage(responseText.String()))
 
 		log.Infof(
-			"Openai %s call took %s with %d sub messages (%d tokens). Message: '%s'. Response: '%s'",
+			"Openai %s call took %s with %d sub messages (%d tokens).",
 			c.cfg.Model,
 			util.FormatDuration(time.Since(startTime)),
 			len(messages),
 			inputTokens,
-			inputText,
-			responseText.String(),
 		)
+		if c.cfg.LogTexts {
+			log.Infof(
+				"Openai texts. Input: '%s'. Response: '%s'",
+				inputText,
+				responseText.String(),
+			)
+		}
 	}()
 }
 

command/openai/config.go

@@ -13,6 +13,8 @@ type Config struct {
 	InitialSystemMessage string  `mapstructure:"initial_system_message"`
 	Model                string  `mapstructure:"model"`
 	Temperature          float32 `mapstructure:"temperature"`
+	Seed                 string  `mapstructure:"seed"`
+	MaxTokens            int     `mapstructure:"max_tokens"`
 
 	// number of thread messages stored which are used as a context for further requests
 	HistorySize int `mapstructure:"history_size"`
@@ -22,6 +24,9 @@ type Config struct {
 
 	// maximum update frequency of slack messages when "stream" is active
 	UpdateInterval time.Duration `mapstructure:"update_interval"`
+
+	// log all input+output text to the logger. This could include personal information, therefore disabled by default!
+	LogTexts bool `mapstructure:"log_texts"`
 }
 
 // IsEnabled checks if token is set

command/openai/tokens.go

@@ -9,15 +9,17 @@ import (
 var maxTokens = map[string]int{
 	"gpt-4":                8192,
 	"gpt-4-32k":            32768,
-	"gpt-3.5-turbo-16k":    16385,
-	"gpt-3.5-turbo":        4096,
 	"gpt-4-1106-preview":   128000,
 	"gpt-4-vision-preview": 128000,
+	"gpt-3.5-turbo-16k":    16385,
+	"gpt-3.5-turbo":        4096,
 	"dummy-test":           100, // just for testing
 }
 
 var modelDateRe = regexp.MustCompile(`-\d{4}`)
 
+// truncateMessages will truncate the messages to fit into the max tokens limit of the model
+// we always try to keep the last message, so we will truncate the first messages
 func truncateMessages(model string, inputMessages []ChatMessage) ([]ChatMessage, int, int) {
 	outputMessages := make([]ChatMessage, 0, len(inputMessages))
 
@@ -47,12 +49,12 @@ func getMaxTokensForModel(model string) int {
 		return getMaxTokensForModel(modelDateRe.ReplaceAllString(model, ""))
 	}
 
-	// we need some default
-	return 4000
+	// we need some default, keep it high, as new models will most likely support more tokens
+	return 128000
 }
 
+// to lower the dependency to heavy external libs we use the rule of thumbs which is totally fine here
+// https://platform.openai.com/tokenizer
 func estimateTokensForMessage(message string) int {
-	// to lower the dependency to heavy external libs we use the rule of thumbs which is totally fine here
-	// https://platform.openai.com/tokenizer
 	return len(message) / 4
 }

command/openai/tokens_test.go

@@ -11,8 +11,8 @@ func TestModels(t *testing.T) {
 		input    string
 		expected int
 	}{
-		{"", 4000},
-		{"jolo", 4000},
+		{"", 128000},
+		{"jolo", 128000},
 		{"gpt-4", 8192},
 		{"gpt-4-0613", 8192},
 		{"gpt-4-32k-0613", 32768},

readme.md

@@ -49,7 +49,7 @@ oauth_config:
       - reactions:read
       - reactions:write
       - users:read
-      - files:write
+      - files:read
 settings:
   event_subscriptions:
     bot_events:
@@ -302,6 +302,7 @@ openai:
   update_interval: '3s' # fewer Slack messages update during generation
   model: gpt-3.5-turbo
   temperature: 0.8
+  log_texts: true # opt in: log all input/output text to the log
 ```
 
 When using the "openai XXX" command within a existing thread, the previous messages are used as context for further calls.