Files
2026-07-13 12:49:10 +08:00

1938 lines
59 KiB
Go

/*
* Copyright 2024 CloudWeGo Authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package schema
import (
"context"
"fmt"
"io"
"reflect"
"sort"
"strings"
"sync"
"text/template"
"github.com/nikolalohinski/gonja"
"github.com/nikolalohinski/gonja/config"
"github.com/nikolalohinski/gonja/exec"
"github.com/nikolalohinski/gonja/nodes"
"github.com/nikolalohinski/gonja/parser"
"github.com/slongfield/pyfmt"
"github.com/cloudwego/eino/internal"
"github.com/cloudwego/eino/internal/generic"
)
func init() {
internal.RegisterStreamChunkConcatFunc(ConcatMessages)
internal.RegisterStreamChunkConcatFunc(ConcatMessageArray)
internal.RegisterStreamChunkConcatFunc(ConcatAgenticMessages)
internal.RegisterStreamChunkConcatFunc(ConcatAgenticMessagesArray)
internal.RegisterStreamChunkConcatFunc(ConcatToolResults)
}
func buildConcatGenericArray[T any](f func([]*T) (*T, error)) func([][]*T) ([]*T, error) {
return func(mas [][]*T) ([]*T, error) {
arrayLen := len(mas[0])
ret := make([]*T, arrayLen)
slicesToConcat := make([][]*T, arrayLen)
for _, ma := range mas {
if len(ma) != arrayLen {
return nil, fmt.Errorf("unexpected array length. "+
"Got %d, expected %d", len(ma), arrayLen)
}
for i := 0; i < arrayLen; i++ {
m := ma[i]
if m != nil {
slicesToConcat[i] = append(slicesToConcat[i], m)
}
}
}
for i, slice := range slicesToConcat {
if len(slice) == 0 {
ret[i] = nil
} else if len(slice) == 1 {
ret[i] = slice[0]
} else {
cm, err := f(slice)
if err != nil {
return nil, err
}
ret[i] = cm
}
}
return ret, nil
}
}
// ConcatMessageArray merges aligned slices of messages into a single slice,
// concatenating messages at the same index across the input arrays.
func ConcatMessageArray(mas [][]*Message) ([]*Message, error) {
return buildConcatGenericArray[Message](ConcatMessages)(mas)
}
// FormatType used by MessageTemplate.Format
type FormatType uint8
const (
// FString Supported by pyfmt(github.com/slongfield/pyfmt), which is an implementation of https://peps.python.org/pep-3101/.
FString FormatType = 0
// GoTemplate https://pkg.go.dev/text/template.
GoTemplate FormatType = 1
// Jinja2 Supported by gonja(github.com/nikolalohinski/gonja), which is a implementation of https://jinja.palletsprojects.com/en/3.1.x/templates/.
Jinja2 FormatType = 2
)
// RoleType is the type of the role of a message.
type RoleType string
const (
// Assistant is the role of an assistant, means the message is returned by ChatModel.
Assistant RoleType = "assistant"
// User is the role of a user, means the message is a user message.
User RoleType = "user"
// System is the role of a system, means the message is a system message.
System RoleType = "system"
// Tool is the role of a tool, means the message is a tool call output.
Tool RoleType = "tool"
)
// FunctionCall is the function call in a message.
// It's used in Assistant Message.
type FunctionCall struct {
// Name is the name of the function to call, it can be used to identify the specific function.
Name string `json:"name,omitempty"`
// Arguments is the arguments to call the function with, in JSON format.
Arguments string `json:"arguments,omitempty"`
}
// ToolCall is the tool call in a message.
// It's used in Assistant Message when there are tool calls should be made.
type ToolCall struct {
// Index is used when there are multiple tool calls in a message.
// In stream mode, it's used to identify the chunk of the tool call for merging.
Index *int `json:"index,omitempty"`
// ID is the id of the tool call, it can be used to identify the specific tool call.
ID string `json:"id"`
// Type is the type of the tool call, default is "function".
Type string `json:"type"`
// Function is the function call to be made.
Function FunctionCall `json:"function"`
// Extra is used to store extra information for the tool call.
Extra map[string]any `json:"extra,omitempty"`
}
// ImageURLDetail is the detail of the image url.
type ImageURLDetail string
const (
// ImageURLDetailHigh means the high quality image url.
ImageURLDetailHigh ImageURLDetail = "high"
// ImageURLDetailLow means the low quality image url.
ImageURLDetailLow ImageURLDetail = "low"
// ImageURLDetailAuto means the auto quality image url.
ImageURLDetailAuto ImageURLDetail = "auto"
)
// MessagePartCommon represents the common abstract components for input and output of multi-modal types.
type MessagePartCommon struct {
// URL is primarily used for HTTP or HTTPS access links.
// For data in the format 'data:[<mediatype>][;base64],<data>' (the 'data' URL Schema of RFC-2397 (https://www.rfc-editor.org/rfc/rfc2397)),
// it is recommended to use Base64Data and MIMEType fields separately instead.
URL *string `json:"url,omitempty"`
// Base64Data represents the binary data in Base64 encoded string format.
Base64Data *string `json:"base64data,omitempty"`
// MIMEType is the mime type , eg."image/png",""audio/wav" etc.
MIMEType string `json:"mime_type,omitempty"`
// Deprecated: Use MessageOutputPart.Extra or MessageInputPart.Extra to set additional metadata instead.
Extra map[string]any `json:"extra,omitempty"`
}
// MessageInputImage is used to represent an image part in message.
// Choose either URL or Base64Data.
type MessageInputImage struct {
MessagePartCommon
// Detail is the quality of the image url.
Detail ImageURLDetail `json:"detail,omitempty"`
}
// MessageInputAudio is used to represent an audio part in message.
// Choose either URL or Base64Data.
type MessageInputAudio struct {
MessagePartCommon
}
// MessageInputVideo is used to represent a video part in message.
// Choose either URL or Base64Data.
type MessageInputVideo struct {
MessagePartCommon
}
// MessageInputFile is used to represent a file part in message.
// Choose either URL or Base64Data.
type MessageInputFile struct {
MessagePartCommon
// Name represents the filename.
// Optional.
Name string `json:"name,omitempty"`
}
// MessageInputPart represents the input part of message.
type MessageInputPart struct {
Type ChatMessagePartType `json:"type"`
Text string `json:"text,omitempty"`
// Image is the image input of the part, it's used when Type is "image_url".
Image *MessageInputImage `json:"image,omitempty"`
// Audio is the audio input of the part, it's used when Type is "audio_url".
Audio *MessageInputAudio `json:"audio,omitempty"`
// Video is the video input of the part, it's used when Type is "video_url".
Video *MessageInputVideo `json:"video,omitempty"`
// File is the file input of the part, it's used when Type is "file_url".
File *MessageInputFile `json:"file,omitempty"`
// ToolSearchResult holds the result of a tool search request, containing the matched tool names and their definitions.
ToolSearchResult *ToolSearchResult `json:"tool_search_result,omitempty"`
// Extra is used to store extra information.
Extra map[string]any `json:"extra,omitempty"`
}
// MessageOutputImage is used to represent an image part in message.
type MessageOutputImage struct {
MessagePartCommon
}
// MessageOutputAudio is used to represent an audio part in message.
type MessageOutputAudio struct {
MessagePartCommon
}
// MessageOutputVideo is used to represent a video part in message.
type MessageOutputVideo struct {
MessagePartCommon
}
// MessageOutputReasoning represents the reasoning content generated by reasoning models.
// Some models produce reasoning steps before generating the final response.
// This struct captures that reasoning output.
type MessageOutputReasoning struct {
// Text is either the thought summary or the raw reasoning text itself.
Text string `json:"text,omitempty"`
// Signature contains encrypted reasoning tokens.
// Required by some models when passing reasoning context back in subsequent requests.
Signature string `json:"signature,omitempty"`
}
// MessageStreamingMeta contains metadata for streaming responses.
// It is used to track position of part when the model outputs multiple parts in a single response.
type MessageStreamingMeta struct {
// Index specifies the index position of this part in the final response.
// This is useful for reassembling multiple reasoning/content parts in correct order.
Index int `json:"index,omitempty"`
}
// MessageOutputPart represents a part of an assistant-generated message.
// It can contain text, or multimedia content like images, audio, or video.
type MessageOutputPart struct {
// Type is the type of the part, e.g. "text", "image_url", "audio_url", "video_url".
Type ChatMessagePartType `json:"type"`
// Text is the text of the part, it's used when Type is "text".
Text string `json:"text,omitempty"`
// Image is the image output of the part, used when Type is ChatMessagePartTypeImageURL.
Image *MessageOutputImage `json:"image,omitempty"`
// Audio is the audio output of the part, used when Type is ChatMessagePartTypeAudioURL.
Audio *MessageOutputAudio `json:"audio,omitempty"`
// Video is the video output of the part, used when Type is ChatMessagePartTypeVideoURL.
Video *MessageOutputVideo `json:"video,omitempty"`
// Reasoning contains the reasoning content generated by the model.
// Used when Type is ChatMessagePartTypeReasoning.
Reasoning *MessageOutputReasoning `json:"reasoning,omitempty"`
// Extra is used to store extra information.
Extra map[string]any `json:"extra,omitempty"`
// StreamingMeta contains metadata for streaming responses.
// This field is typically used at runtime and not serialized.
StreamingMeta *MessageStreamingMeta `json:"-"`
}
// Deprecated: This struct is deprecated as the MultiContent field is deprecated.
// For the image input part of the model, use MessageInputImage.
// For the image output part of the model, use MessageOutputImage.
// Choose either URL or URI.
// If your model implementation supports it, URL could embed inline image data
// as defined in RFC-2397.
type ChatMessageImageURL struct {
// URL can either be a traditional URL or a special URL conforming to RFC-2397 (https://www.rfc-editor.org/rfc/rfc2397).
// double check with model implementations for detailed instructions on how to use this.
URL string `json:"url,omitempty"`
URI string `json:"uri,omitempty"`
// Detail is the quality of the image url.
Detail ImageURLDetail `json:"detail,omitempty"`
// MIMEType is the mime type of the image, eg. "image/png".
MIMEType string `json:"mime_type,omitempty"`
// Extra is used to store extra information for the image url.
Extra map[string]any `json:"extra,omitempty"`
}
// ChatMessagePartType is the type of the part in a chat message.
type ChatMessagePartType string
const (
// ChatMessagePartTypeText means the part is a text.
ChatMessagePartTypeText ChatMessagePartType = "text"
// ChatMessagePartTypeImageURL means the part is an image url.
ChatMessagePartTypeImageURL ChatMessagePartType = "image_url"
// ChatMessagePartTypeAudioURL means the part is an audio url.
ChatMessagePartTypeAudioURL ChatMessagePartType = "audio_url"
// ChatMessagePartTypeVideoURL means the part is a video url.
ChatMessagePartTypeVideoURL ChatMessagePartType = "video_url"
// ChatMessagePartTypeFileURL means the part is a file url.
ChatMessagePartTypeFileURL ChatMessagePartType = "file_url"
// ChatMessagePartTypeReasoning means the part is a reasoning block.
ChatMessagePartTypeReasoning ChatMessagePartType = "reasoning"
// ChatMessagePartTypeToolSearchResult means the part contains tool search results.
ChatMessagePartTypeToolSearchResult ChatMessagePartType = "tool_search_result"
)
// Deprecated: This struct is deprecated as the MultiContent field is deprecated.
// For the audio input part of the model, use MessageInputAudio.
// For the audio output part of the model, use MessageOutputAudio.
// Choose either URL or URI.
// If supported, URL may embed inline audio data per RFC-2397.
type ChatMessageAudioURL struct {
// URL can either be a traditional URL or a special URL conforming to RFC-2397 (https://www.rfc-editor.org/rfc/rfc2397).
// double check with model implementations for detailed instructions on how to use this.
URL string `json:"url,omitempty"`
URI string `json:"uri,omitempty"`
// MIMEType is the mime type of the audio, eg. "audio/wav" or "audio/ogg".
MIMEType string `json:"mime_type,omitempty"`
// Extra is used to store extra information for the audio url.
Extra map[string]any `json:"extra,omitempty"`
}
// Deprecated: This struct is deprecated as the MultiContent field is deprecated.
// For the video input part of the model, use MessageInputVideo.
// For the video output part of the model, use MessageOutputVideo.
// Choose either URL or URI.
// If supported, URL may embed inline video data per RFC-2397.
type ChatMessageVideoURL struct {
// URL can either be a traditional URL or a special URL conforming to RFC-2397 (https://www.rfc-editor.org/rfc/rfc2397).
// double check with model implementations for detailed instructions on how to use this.
URL string `json:"url,omitempty"`
URI string `json:"uri,omitempty"`
// MIMEType is the mime type of the video, eg. "video/mp4".
MIMEType string `json:"mime_type,omitempty"`
// Extra is used to store extra information for the video url.
Extra map[string]any `json:"extra,omitempty"`
}
// Deprecated: This struct is deprecated as the MultiContent field is deprecated.
// For the file input part of the model, use MessageInputFile.
// Choose either URL or URI.
type ChatMessageFileURL struct {
URL string `json:"url,omitempty"`
URI string `json:"uri,omitempty"`
// MIMEType is the mime type of the file, eg. "application/pdf", "text/plain".
MIMEType string `json:"mime_type,omitempty"`
// Name is the name of the file.
Name string `json:"name,omitempty"`
// Extra is used to store extra information for the file url.
Extra map[string]any `json:"extra,omitempty"`
}
// Deprecated: This struct is deprecated as the MultiContent field is deprecated.
// For model input, use MessageInputPart. For model output, use MessageOutputPart.
type ChatMessagePart struct {
// Type is the type of the part, eg. "text", "image_url", "audio_url", "video_url", "file_url".
Type ChatMessagePartType `json:"type,omitempty"`
// Text is the text of the part, it's used when Type is "text".
Text string `json:"text,omitempty"`
// ImageURL is the image url of the part, it's used when Type is "image_url".
ImageURL *ChatMessageImageURL `json:"image_url,omitempty"`
// AudioURL is the audio url of the part, it's used when Type is "audio_url".
AudioURL *ChatMessageAudioURL `json:"audio_url,omitempty"`
// VideoURL is the video url of the part, it's used when Type is "video_url".
VideoURL *ChatMessageVideoURL `json:"video_url,omitempty"`
// FileURL is the file url of the part, it's used when Type is "file_url".
FileURL *ChatMessageFileURL `json:"file_url,omitempty"`
}
// LogProbs is the top-level structure containing the log probability information.
type LogProbs struct {
// Content is a list of message content tokens with log probability information.
Content []LogProb `json:"content"`
}
// LogProb represents the probability information for a token.
type LogProb struct {
// Token represents the text of the token, which is a contiguous sequence of characters
// (e.g., a word, part of a word, or punctuation) as understood by the tokenization process used by the language model.
Token string `json:"token"`
// LogProb is the log probability of this token, if it is within the top 20 most likely tokens.
// Otherwise, the value `-9999.0` is used to signify that the token is very unlikely.
LogProb float64 `json:"logprob"`
// Bytes is a list of integers representing the UTF-8 bytes representation of the token.
// Useful in instances where characters are represented by multiple tokens and
// their byte representations must be combined to generate the correct text
// representation. Can be `null` if there is no bytes representation for the token.
Bytes []int64 `json:"bytes,omitempty"` // Omitting the field if it is null
// TopLogProbs is a list of the most likely tokens and their log probability, at this token position.
// In rare cases, there may be fewer than the number of requested top_logprobs returned.
TopLogProbs []TopLogProb `json:"top_logprobs"`
}
// TopLogProb describes a likely token and its log probability at a position.
type TopLogProb struct {
// Token represents the text of the token, which is a contiguous sequence of characters
// (e.g., a word, part of a word, or punctuation) as understood by the tokenization process used by the language model.
Token string `json:"token"`
// LogProb is the log probability of this token, if it is within the top 20 most likely tokens.
// Otherwise, the value `-9999.0` is used to signify that the token is very unlikely.
LogProb float64 `json:"logprob"`
// Bytes is a list of integers representing the UTF-8 bytes representation of the token.
// Useful in instances where characters are represented by multiple tokens and
// their byte representations must be combined to generate the correct text
// representation. Can be `null` if there is no bytes representation for the token.
Bytes []int64 `json:"bytes,omitempty"`
}
// ResponseMeta collects meta information about a chat response.
type ResponseMeta struct {
// FinishReason is the reason why the chat response is finished.
// It's usually "stop", "length", "tool_calls", "content_filter", "null". This is defined by chat model implementation.
FinishReason string `json:"finish_reason,omitempty"`
// Usage is the token usage of the chat response, whether usage exists depends on whether the chat model implementation returns.
Usage *TokenUsage `json:"usage,omitempty"`
// LogProbs is Log probability information.
LogProbs *LogProbs `json:"logprobs,omitempty"`
}
// Message denotes the data structure for model input and output, originating from either user input or model return.
// It supports both text-only and multimodal content.
//
// For text-only input from a user, use the Content field:
//
// &schema.Message{
// Role: schema.User,
// Content: "What is the capital of France?",
// }
//
// For multimodal input from a user, use the UserInputMultiContent field.
// This allows combining text with other media like images:
//
// &schema.Message{
// Role: schema.User,
// UserInputMultiContent: []schema.MessageInputPart{
// {Type: schema.ChatMessagePartTypeText, Text: "What is in this image?"},
// {Type: schema.ChatMessagePartTypeImageURL, Image: &schema.MessageInputImage{
// MessagePartCommon: schema.MessagePartCommon{
// URL: toPtr("https://example.com/cat.jpg"),
// },
// Detail: schema.ImageURLDetailHigh,
// }},
// },
// }
//
// When the model returns multimodal content, it is available in the AssistantGenMultiContent field:
//
// &schema.Message{
// Role: schema.Assistant,
// AssistantGenMultiContent: []schema.MessageOutputPart{
// {Type: schema.ChatMessagePartTypeText, Text: "Here is the generated image:"},
// {Type: schema.ChatMessagePartTypeImage, Image: &schema.MessageOutputImage{
// MessagePartCommon: schema.MessagePartCommon{
// Base64Data: toPtr("base64_image_binary"),
// MIMEType: "image/png",
// },
// }},
// },
// }
type Message struct {
Role RoleType `json:"role"`
// Content is for user text input and model text output.
Content string `json:"content"`
// if MultiContent is not empty, use this instead of Content
// if MultiContent is empty, use Content
// Deprecated: Use UserInputMultiContent for user multimodal inputs and AssistantGenMultiContent for model multimodal outputs.
MultiContent []ChatMessagePart `json:"multi_content,omitempty"`
// UserInputMultiContent passes multimodal content provided by the user to the model.
UserInputMultiContent []MessageInputPart `json:"user_input_multi_content,omitempty"`
// AssistantGenMultiContent is for receiving multimodal output from the model.
AssistantGenMultiContent []MessageOutputPart `json:"assistant_output_multi_content,omitempty"`
Name string `json:"name,omitempty"`
// only for AssistantMessage
ToolCalls []ToolCall `json:"tool_calls,omitempty"`
// only for ToolMessage
ToolCallID string `json:"tool_call_id,omitempty"`
// only for ToolMessage
ToolName string `json:"tool_name,omitempty"`
ResponseMeta *ResponseMeta `json:"response_meta,omitempty"`
// ReasoningContent is the thinking process of the model, which will be included when the model returns reasoning content.
ReasoningContent string `json:"reasoning_content,omitempty"`
// customized information for model implementation
Extra map[string]any `json:"extra,omitempty"`
}
// TokenUsage Represents the token usage of chat model request.
type TokenUsage struct {
// PromptTokens is the number of prompt tokens, including all the input tokens of this request.
PromptTokens int `json:"prompt_tokens"`
// PromptTokenDetails is a breakdown of the prompt tokens.
PromptTokenDetails PromptTokenDetails `json:"prompt_token_details"`
// CompletionTokens is the number of completion tokens.
CompletionTokens int `json:"completion_tokens"`
// TotalTokens is the total number of tokens.
TotalTokens int `json:"total_tokens"`
// CompletionTokensDetails is breakdown of completion tokens.
CompletionTokensDetails CompletionTokensDetails `json:"completion_token_details"`
}
type CompletionTokensDetails struct {
// ReasoningTokens tokens generated by the model for reasoning.
// This is currently supported by OpenAI, Gemini, ARK and Qwen chat models.
// For other models, this field will be 0.
ReasoningTokens int `json:"reasoning_tokens,omitempty"`
}
// PromptTokenDetails provides a breakdown of prompt token usage.
type PromptTokenDetails struct {
// Cached tokens present in the prompt.
CachedTokens int `json:"cached_tokens"`
}
var _ MessagesTemplate = &Message{}
var _ MessagesTemplate = MessagesPlaceholder("", false)
// MessagesTemplate is the interface for messages template.
// It's used to render a template to a list of messages.
// e.g.
//
// chatTemplate := prompt.FromMessages(
// schema.SystemMessage("you are an eino helper"),
// schema.MessagesPlaceholder("history", false), // <= this will use the value of "history" in params
// )
// msgs, err := chatTemplate.Format(ctx, params)
type MessagesTemplate interface {
Format(ctx context.Context, vs map[string]any, formatType FormatType) ([]*Message, error)
}
type messagesPlaceholder struct {
key string
optional bool
}
// MessagesPlaceholder can render a placeholder to a list of messages in params.
// e.g.
//
// placeholder := MessagesPlaceholder("history", false)
// params := map[string]any{
// "history": []*schema.Message{{Role: "user", Content: "what is eino?"}, {Role: "assistant", Content: "eino is a great framework to build llm apps"}},
// "query": "how to use eino?",
// }
// chatTemplate := chatTpl := prompt.FromMessages(
// schema.SystemMessage("you are eino helper"),
// schema.MessagesPlaceholder("history", false), // <= this will use the value of "history" in params
// )
// msgs, err := chatTemplate.Format(ctx, params)
func MessagesPlaceholder(key string, optional bool) MessagesTemplate {
return &messagesPlaceholder{
key: key,
optional: optional,
}
}
// Format just return the messages of specified key.
// because it's a placeholder.
// e.g.
//
// placeholder := MessagesPlaceholder("history", false)
// params := map[string]any{
// "history": []*schema.Message{{Role: "user", Content: "what is eino?"}, {Role: "assistant", Content: "eino is a great freamwork to build llm apps"}},
// "query": "how to use eino?",
// }
// msgs, err := placeholder.Format(ctx, params) // <= this will return the value of "history" in params
func (p *messagesPlaceholder) Format(_ context.Context, vs map[string]any, _ FormatType) ([]*Message, error) {
v, ok := vs[p.key]
if !ok {
if p.optional {
return []*Message{}, nil
}
return nil, fmt.Errorf("message placeholder format: %s not found", p.key)
}
msgs, ok := v.([]*Message)
if !ok {
return nil, fmt.Errorf("only messages can be used to format message placeholder, key: %v, actual type: %v", p.key, reflect.TypeOf(v))
}
return msgs, nil
}
func formatContent(content string, vs map[string]any, formatType FormatType) (string, error) {
switch formatType {
case FString:
return pyfmt.Fmt(content, vs)
case GoTemplate:
parsedTmpl, err := template.New("template").
Option("missingkey=error").
Parse(content)
if err != nil {
return "", err
}
sb := new(strings.Builder)
err = parsedTmpl.Execute(sb, vs)
if err != nil {
return "", err
}
return sb.String(), nil
case Jinja2:
env, err := getJinjaEnv()
if err != nil {
return "", err
}
tpl, err := env.FromString(content)
if err != nil {
return "", err
}
out, err := tpl.Execute(vs)
if err != nil {
return "", err
}
return out, nil
default:
return "", fmt.Errorf("unknown format type: %v", formatType)
}
}
// Format returns the messages after rendering by the given formatType.
// e.g.
//
// msg := schema.UserMessage("hello world, {name}")
// msgs, err := msg.Format(ctx, map[string]any{"name": "eino"}, schema.FString) // <= this will render the content of msg by pyfmt
// // msgs[0].Content will be "hello world, eino"
func (m *Message) Format(_ context.Context, vs map[string]any, formatType FormatType) ([]*Message, error) {
c, err := formatContent(m.Content, vs, formatType)
if err != nil {
return nil, err
}
copied := *m
copied.Content = c
if len(m.MultiContent) > 0 {
copied.MultiContent, err = formatMultiContent(m.MultiContent, vs, formatType)
if err != nil {
return nil, err
}
}
if len(m.UserInputMultiContent) > 0 {
copied.UserInputMultiContent, err = formatUserInputMultiContent(m.UserInputMultiContent, vs, formatType)
if err != nil {
return nil, err
}
}
return []*Message{&copied}, nil
}
func formatMultiContent(multiContent []ChatMessagePart, vs map[string]any, formatType FormatType) ([]ChatMessagePart, error) {
copiedMC := make([]ChatMessagePart, len(multiContent))
copy(copiedMC, multiContent)
for i, mc := range copiedMC {
switch mc.Type {
case ChatMessagePartTypeText:
nmc, err := formatContent(mc.Text, vs, formatType)
if err != nil {
return nil, err
}
copiedMC[i].Text = nmc
case ChatMessagePartTypeImageURL:
if mc.ImageURL == nil {
continue
}
url, err := formatContent(mc.ImageURL.URL, vs, formatType)
if err != nil {
return nil, err
}
copiedMC[i].ImageURL.URL = url
case ChatMessagePartTypeAudioURL:
if mc.AudioURL == nil {
continue
}
url, err := formatContent(mc.AudioURL.URL, vs, formatType)
if err != nil {
return nil, err
}
copiedMC[i].AudioURL.URL = url
case ChatMessagePartTypeVideoURL:
if mc.VideoURL == nil {
continue
}
url, err := formatContent(mc.VideoURL.URL, vs, formatType)
if err != nil {
return nil, err
}
copiedMC[i].VideoURL.URL = url
case ChatMessagePartTypeFileURL:
if mc.FileURL == nil {
continue
}
url, err := formatContent(mc.FileURL.URL, vs, formatType)
if err != nil {
return nil, err
}
copiedMC[i].FileURL.URL = url
}
}
return copiedMC, nil
}
func formatUserInputMultiContent(userInputMultiContent []MessageInputPart, vs map[string]any, formatType FormatType) ([]MessageInputPart, error) {
copiedUIMC := make([]MessageInputPart, len(userInputMultiContent))
copy(copiedUIMC, userInputMultiContent)
for i, uimc := range copiedUIMC {
switch uimc.Type {
case ChatMessagePartTypeText:
text, err := formatContent(uimc.Text, vs, formatType)
if err != nil {
return nil, err
}
copiedUIMC[i].Text = text
case ChatMessagePartTypeImageURL:
if uimc.Image == nil {
continue
}
if uimc.Image.URL != nil && *uimc.Image.URL != "" {
url, err := formatContent(*uimc.Image.URL, vs, formatType)
if err != nil {
return nil, err
}
copiedUIMC[i].Image.URL = &url
}
if uimc.Image.Base64Data != nil && *uimc.Image.Base64Data != "" {
base64data, err := formatContent(*uimc.Image.Base64Data, vs, formatType)
if err != nil {
return nil, err
}
copiedUIMC[i].Image.Base64Data = &base64data
}
case ChatMessagePartTypeAudioURL:
if uimc.Audio == nil {
continue
}
if uimc.Audio.URL != nil && *uimc.Audio.URL != "" {
url, err := formatContent(*uimc.Audio.URL, vs, formatType)
if err != nil {
return nil, err
}
copiedUIMC[i].Audio.URL = &url
}
if uimc.Audio.Base64Data != nil && *uimc.Audio.Base64Data != "" {
base64data, err := formatContent(*uimc.Audio.Base64Data, vs, formatType)
if err != nil {
return nil, err
}
copiedUIMC[i].Audio.Base64Data = &base64data
}
case ChatMessagePartTypeVideoURL:
if uimc.Video == nil {
continue
}
if uimc.Video.URL != nil && *uimc.Video.URL != "" {
url, err := formatContent(*uimc.Video.URL, vs, formatType)
if err != nil {
return nil, err
}
copiedUIMC[i].Video.URL = &url
}
if uimc.Video.Base64Data != nil && *uimc.Video.Base64Data != "" {
base64data, err := formatContent(*uimc.Video.Base64Data, vs, formatType)
if err != nil {
return nil, err
}
copiedUIMC[i].Video.Base64Data = &base64data
}
case ChatMessagePartTypeFileURL:
if uimc.File == nil {
continue
}
if uimc.File.URL != nil && *uimc.File.URL != "" {
url, err := formatContent(*uimc.File.URL, vs, formatType)
if err != nil {
return nil, err
}
copiedUIMC[i].File.URL = &url
}
if uimc.File.Base64Data != nil && *uimc.File.Base64Data != "" {
base64data, err := formatContent(*uimc.File.Base64Data, vs, formatType)
if err != nil {
return nil, err
}
copiedUIMC[i].File.Base64Data = &base64data
}
}
}
return copiedUIMC, nil
}
// String returns the string representation of the message.
// e.g.
//
// msg := schema.UserMessage("hello world")
// fmt.Println(msg.String()) // Output will be: `user: hello world``
//
// msg := schema.Message{
// Role: schema.Tool,
// Content: "{...}",
// ToolCallID: "callxxxx"
// }
// fmt.Println(msg.String())
// Output will be:
// tool: {...}
// call_id: callxxxx
func (m *Message) String() string {
sb := &strings.Builder{}
sb.WriteString(fmt.Sprintf("%s: %s", m.Role, m.Content))
if len(m.UserInputMultiContent) > 0 {
sb.WriteString("\nuser_input_multi_content:")
for i, part := range m.UserInputMultiContent {
sb.WriteString(fmt.Sprintf("\n [%d] %s", i, formatInputPart(part)))
}
}
if len(m.AssistantGenMultiContent) > 0 {
sb.WriteString("\nassistant_gen_multi_content:")
for i, part := range m.AssistantGenMultiContent {
sb.WriteString(fmt.Sprintf("\n [%d] %s", i, formatOutputPart(part)))
}
}
if len(m.MultiContent) > 0 {
sb.WriteString("\nmulti_content:")
for i, part := range m.MultiContent {
sb.WriteString(fmt.Sprintf("\n [%d] %s", i, formatChatMessagePart(part)))
}
}
if len(m.ReasoningContent) > 0 {
sb.WriteString("\nreasoning content:\n")
sb.WriteString(m.ReasoningContent)
}
if len(m.ToolCalls) > 0 {
sb.WriteString("\ntool_calls:\n")
for _, tc := range m.ToolCalls {
if tc.Index != nil {
sb.WriteString(fmt.Sprintf("index[%d]:", *tc.Index))
}
sb.WriteString(fmt.Sprintf("%+v\n", tc))
}
}
if m.ToolCallID != "" {
sb.WriteString(fmt.Sprintf("\ntool_call_id: %s", m.ToolCallID))
}
if m.ToolName != "" {
sb.WriteString(fmt.Sprintf("\ntool_call_name: %s", m.ToolName))
}
if m.ResponseMeta != nil {
sb.WriteString(fmt.Sprintf("\nfinish_reason: %s", m.ResponseMeta.FinishReason))
if m.ResponseMeta.Usage != nil {
sb.WriteString(fmt.Sprintf("\nusage: %v", m.ResponseMeta.Usage))
}
}
return sb.String()
}
func formatInputPart(part MessageInputPart) string {
switch part.Type {
case ChatMessagePartTypeText:
return fmt.Sprintf("text: %s", part.Text)
case ChatMessagePartTypeImageURL:
return fmt.Sprintf("image: %s", formatMessageInputMedia(part.Image))
case ChatMessagePartTypeAudioURL:
return fmt.Sprintf("audio: %s", formatMessageInputMedia(part.Audio))
case ChatMessagePartTypeVideoURL:
return fmt.Sprintf("video: %s", formatMessageInputMedia(part.Video))
case ChatMessagePartTypeFileURL:
return fmt.Sprintf("file: %s", formatMessageInputFile(part.File))
default:
return fmt.Sprintf("unknown type: %s", part.Type)
}
}
func formatMessageInputMedia[T MessageInputImage | MessageInputAudio | MessageInputVideo](media *T) string {
if media == nil {
return "<nil>"
}
var parts []string
switch v := any(media).(type) {
case *MessageInputImage:
if v.URL != nil {
parts = append(parts, fmt.Sprintf("url=%s", *v.URL))
}
if v.Base64Data != nil {
parts = append(parts, fmt.Sprintf("base64[%d bytes]", len(*v.Base64Data)))
}
if v.MIMEType != "" {
parts = append(parts, fmt.Sprintf("mime=%s", v.MIMEType))
}
if v.Detail != "" {
parts = append(parts, fmt.Sprintf("detail=%s", v.Detail))
}
if len(v.Extra) > 0 {
parts = append(parts, fmt.Sprintf("extra=%v", v.Extra))
}
case *MessageInputAudio:
if v.URL != nil {
parts = append(parts, fmt.Sprintf("url=%s", *v.URL))
}
if v.Base64Data != nil {
parts = append(parts, fmt.Sprintf("base64[%d bytes]", len(*v.Base64Data)))
}
if v.MIMEType != "" {
parts = append(parts, fmt.Sprintf("mime=%s", v.MIMEType))
}
if len(v.Extra) > 0 {
parts = append(parts, fmt.Sprintf("extra=%v", v.Extra))
}
case *MessageInputVideo:
if v.URL != nil {
parts = append(parts, fmt.Sprintf("url=%s", *v.URL))
}
if v.Base64Data != nil {
parts = append(parts, fmt.Sprintf("base64[%d bytes]", len(*v.Base64Data)))
}
if v.MIMEType != "" {
parts = append(parts, fmt.Sprintf("mime=%s", v.MIMEType))
}
if len(v.Extra) > 0 {
parts = append(parts, fmt.Sprintf("extra=%v", v.Extra))
}
}
if len(parts) == 0 {
return "<empty>"
}
return strings.Join(parts, ", ")
}
func formatMessageInputFile(file *MessageInputFile) string {
if file == nil {
return "<nil>"
}
var parts []string
if file.URL != nil {
parts = append(parts, fmt.Sprintf("url=%s", *file.URL))
}
if file.Base64Data != nil {
parts = append(parts, fmt.Sprintf("base64[%d bytes]", len(*file.Base64Data)))
}
if file.MIMEType != "" {
parts = append(parts, fmt.Sprintf("mime=%s", file.MIMEType))
}
if file.Name != "" {
parts = append(parts, fmt.Sprintf("name=%s", file.Name))
}
if len(file.Extra) > 0 {
parts = append(parts, fmt.Sprintf("extra=%v", file.Extra))
}
if len(parts) == 0 {
return "<empty>"
}
return strings.Join(parts, ", ")
}
func formatOutputPart(part MessageOutputPart) string {
switch part.Type {
case ChatMessagePartTypeText:
return fmt.Sprintf("text: %s", part.Text)
case ChatMessagePartTypeImageURL:
return fmt.Sprintf("image: %s", formatMessageOutputMedia(part.Image))
case ChatMessagePartTypeAudioURL:
return fmt.Sprintf("audio: %s", formatMessageOutputMedia(part.Audio))
case ChatMessagePartTypeVideoURL:
return fmt.Sprintf("video: %s", formatMessageOutputMedia(part.Video))
default:
return fmt.Sprintf("unknown type: %s", part.Type)
}
}
func formatMessageOutputMedia[T MessageOutputImage | MessageOutputAudio | MessageOutputVideo](media *T) string {
if media == nil {
return "<nil>"
}
var parts []string
switch v := any(media).(type) {
case *MessageOutputImage:
if v.URL != nil {
parts = append(parts, fmt.Sprintf("url=%s", *v.URL))
}
if v.Base64Data != nil {
parts = append(parts, fmt.Sprintf("base64[%d bytes]", len(*v.Base64Data)))
}
if v.MIMEType != "" {
parts = append(parts, fmt.Sprintf("mime=%s", v.MIMEType))
}
if len(v.Extra) > 0 {
parts = append(parts, fmt.Sprintf("extra=%v", v.Extra))
}
case *MessageOutputAudio:
if v.URL != nil {
parts = append(parts, fmt.Sprintf("url=%s", *v.URL))
}
if v.Base64Data != nil {
parts = append(parts, fmt.Sprintf("base64[%d bytes]", len(*v.Base64Data)))
}
if v.MIMEType != "" {
parts = append(parts, fmt.Sprintf("mime=%s", v.MIMEType))
}
if len(v.Extra) > 0 {
parts = append(parts, fmt.Sprintf("extra=%v", v.Extra))
}
case *MessageOutputVideo:
if v.URL != nil {
parts = append(parts, fmt.Sprintf("url=%s", *v.URL))
}
if v.Base64Data != nil {
parts = append(parts, fmt.Sprintf("base64[%d bytes]", len(*v.Base64Data)))
}
if v.MIMEType != "" {
parts = append(parts, fmt.Sprintf("mime=%s", v.MIMEType))
}
if len(v.Extra) > 0 {
parts = append(parts, fmt.Sprintf("extra=%v", v.Extra))
}
}
if len(parts) == 0 {
return "<empty>"
}
return strings.Join(parts, ", ")
}
func formatChatMessagePart(part ChatMessagePart) string {
switch part.Type {
case ChatMessagePartTypeText:
return fmt.Sprintf("text: %s", part.Text)
case ChatMessagePartTypeImageURL:
if part.ImageURL != nil {
return fmt.Sprintf("image_url: %s", part.ImageURL.URL)
}
return "image_url: <nil>"
case ChatMessagePartTypeAudioURL:
if part.AudioURL != nil {
return fmt.Sprintf("audio_url: %s", part.AudioURL.URL)
}
return "audio_url: <nil>"
case ChatMessagePartTypeVideoURL:
if part.VideoURL != nil {
return fmt.Sprintf("video_url: %s", part.VideoURL.URL)
}
return "video_url: <nil>"
case ChatMessagePartTypeFileURL:
if part.FileURL != nil {
return fmt.Sprintf("file_url: %s", part.FileURL.URL)
}
return "file_url: <nil>"
default:
return fmt.Sprintf("unknown type: %s", part.Type)
}
}
// SystemMessage represents a message with Role "system".
func SystemMessage(content string) *Message {
return &Message{
Role: System,
Content: content,
}
}
// AssistantMessage represents a message with Role "assistant".
func AssistantMessage(content string, toolCalls []ToolCall) *Message {
return &Message{
Role: Assistant,
Content: content,
ToolCalls: toolCalls,
}
}
// UserMessage represents a message with Role "user".
func UserMessage(content string) *Message {
return &Message{
Role: User,
Content: content,
}
}
type toolMessageOptions struct {
toolName string
}
// ToolMessageOption defines a option for ToolMessage
type ToolMessageOption func(*toolMessageOptions)
// WithToolName returns a ToolMessageOption that sets the tool call name.
func WithToolName(name string) ToolMessageOption {
return func(o *toolMessageOptions) {
o.toolName = name
}
}
// ToolMessage represents a message with Role "tool".
func ToolMessage(content string, toolCallID string, opts ...ToolMessageOption) *Message {
o := &toolMessageOptions{}
for _, opt := range opts {
opt(o)
}
return &Message{
Role: Tool,
Content: content,
ToolCallID: toolCallID,
ToolName: o.toolName,
}
}
// ConcatToolResults merges multiple ToolResult chunks into a single ToolResult.
// It collects all ToolOutputParts from the input chunks and merges contiguous text parts within each chunk.
//
// Merge rules:
// - Text parts: Contiguous text parts within each chunk are concatenated into a single text part.
// - Non-text parts (image, audio, video, file): These parts are kept as-is without merging.
// Each non-text part type can only appear in one chunk; if the same non-text type appears
// in multiple chunks, an error is returned.
//
// This function is primarily used in streaming scenarios where tool output is delivered
// in multiple chunks that need to be merged into a complete result.
//
// Parameters:
// - chunks: A slice of ToolResult pointers representing sequential chunks from a stream.
// Nil chunks and chunks with empty Parts are safely ignored.
//
// Returns:
// - *ToolResult: The merged ToolResult containing all content from the chunks.
// Returns an empty ToolResult if chunks is empty or all chunks are nil/empty.
// - error: An error if the same non-text part type appears in multiple chunks.
func ConcatToolResults(chunks []*ToolResult) (*ToolResult, error) {
if len(chunks) == 0 {
return &ToolResult{}, nil
}
nonTextPartTypes := make(map[ToolPartType]int)
var allParts []ToolOutputPart
for chunkIdx, chunk := range chunks {
if chunk == nil || len(chunk.Parts) == 0 {
continue
}
for _, part := range chunk.Parts {
if part.Type != ToolPartTypeText {
// This restricts non-text modal content only appear once.
if prevChunkIdx, exists := nonTextPartTypes[part.Type]; exists {
return nil, fmt.Errorf("conflicting %s parts found in chunk %d and chunk %d: "+
"non-text modality parts cannot appear in multiple chunks", part.Type, prevChunkIdx, chunkIdx)
}
nonTextPartTypes[part.Type] = chunkIdx
}
}
allParts = append(allParts, chunk.Parts...)
}
mergedChunkParts, err := concatToolOutputParts(allParts)
if err != nil {
return nil, fmt.Errorf("failed to merge parts: %w", err)
}
return &ToolResult{Parts: mergedChunkParts}, nil
}
func concatToolOutputParts(parts []ToolOutputPart) ([]ToolOutputPart, error) {
if len(parts) == 0 {
return nil, nil
}
groups := groupToolOutputParts(parts)
merged := make([]ToolOutputPart, 0, len(groups))
for _, group := range groups {
if len(group) == 1 {
merged = append(merged, group...)
continue
}
switch group[0].Type {
case ToolPartTypeText:
mergedPart, err := mergeToolTextParts(group)
if err != nil {
return nil, err
}
merged = append(merged, mergedPart)
default:
merged = append(merged, group...)
}
}
return merged, nil
}
func groupToolOutputParts(parts []ToolOutputPart) [][]ToolOutputPart {
groups := make([][]ToolOutputPart, 0)
i := 0
for i < len(parts) {
if parts[i].Type == ToolPartTypeText {
end := i + 1
for end < len(parts) && parts[end].Type == ToolPartTypeText {
end++
}
groups = append(groups, parts[i:end])
i = end
} else {
groups = append(groups, parts[i:i+1])
i++
}
}
return groups
}
func mergeToolTextParts(group []ToolOutputPart) (ToolOutputPart, error) {
var sb strings.Builder
extraList := make([]map[string]any, 0, len(group))
for _, part := range group {
sb.WriteString(part.Text)
if len(part.Extra) > 0 {
extraList = append(extraList, part.Extra)
}
}
var mergedExtra map[string]any
if len(extraList) > 0 {
var err error
mergedExtra, err = concatExtra(extraList)
if err != nil {
return ToolOutputPart{}, fmt.Errorf("failed to concat tool output text part extra: %w", err)
}
}
return ToolOutputPart{
Type: ToolPartTypeText,
Text: sb.String(),
Extra: mergedExtra,
}, nil
}
func concatToolCalls(chunks []ToolCall) ([]ToolCall, error) {
var merged []ToolCall
m := make(map[int][]int)
for i := range chunks {
index := chunks[i].Index
if index == nil {
merged = append(merged, chunks[i])
} else {
m[*index] = append(m[*index], i)
}
}
var args strings.Builder
for k, v := range m {
index := k
toolCall := ToolCall{Index: &index}
if len(v) > 0 {
toolCall = chunks[v[0]]
}
args.Reset()
toolID, toolType, toolName := "", "", "" // these field will output atomically in any chunk
for _, n := range v {
chunk := chunks[n]
if chunk.ID != "" {
if toolID == "" {
toolID = chunk.ID
} else if toolID != chunk.ID {
return nil, fmt.Errorf("cannot concat ToolCalls with different tool id: '%s' '%s'", toolID, chunk.ID)
}
}
if chunk.Type != "" {
if toolType == "" {
toolType = chunk.Type
} else if toolType != chunk.Type {
return nil, fmt.Errorf("cannot concat ToolCalls with different tool type: '%s' '%s'", toolType, chunk.Type)
}
}
if chunk.Function.Name != "" {
if toolName == "" {
toolName = chunk.Function.Name
} else if toolName != chunk.Function.Name {
return nil, fmt.Errorf("cannot concat ToolCalls with different tool name: '%s' '%s'", toolName, chunk.Function.Name)
}
}
if chunk.Function.Arguments != "" {
_, err := args.WriteString(chunk.Function.Arguments)
if err != nil {
return nil, err
}
}
}
toolCall.ID = toolID
toolCall.Type = toolType
toolCall.Function.Name = toolName
toolCall.Function.Arguments = args.String()
merged = append(merged, toolCall)
}
if len(merged) > 1 {
sort.SliceStable(merged, func(i, j int) bool {
iVal, jVal := merged[i].Index, merged[j].Index
if iVal == nil && jVal == nil {
return false
} else if iVal == nil && jVal != nil {
return true
} else if iVal != nil && jVal == nil {
return false
}
return *iVal < *jVal
})
}
return merged, nil
}
func concatAssistantMultiContent(parts []MessageOutputPart) ([]MessageOutputPart, error) {
if len(parts) == 0 {
return parts, nil
}
groups := groupOutputParts(parts)
merged := make([]MessageOutputPart, 0, len(groups))
for _, group := range groups {
mergedPart, err := mergeOutputPartGroup(group)
if err != nil {
return nil, err
}
merged = append(merged, mergedPart)
}
return merged, nil
}
func groupOutputParts(parts []MessageOutputPart) [][]MessageOutputPart {
if len(parts) == 0 {
return nil
}
groups := make([][]MessageOutputPart, 0)
currentGroup := []MessageOutputPart{parts[0]}
for i := 1; i < len(parts); i++ {
if canMergeOutputParts(currentGroup[0], parts[i]) {
currentGroup = append(currentGroup, parts[i])
} else {
groups = append(groups, currentGroup)
currentGroup = []MessageOutputPart{parts[i]}
}
}
groups = append(groups, currentGroup)
return groups
}
func canMergeOutputParts(current, next MessageOutputPart) bool {
if current.Type != next.Type {
return false
}
if !isMergeableOutputPartType(current) {
return false
}
if current.StreamingMeta != nil && next.StreamingMeta != nil {
return current.StreamingMeta.Index == next.StreamingMeta.Index
}
return current.StreamingMeta == nil && next.StreamingMeta == nil
}
func isMergeableOutputPartType(part MessageOutputPart) bool {
switch part.Type {
case ChatMessagePartTypeText, ChatMessagePartTypeReasoning:
return true
case ChatMessagePartTypeAudioURL:
return isBase64MessageOutputAudioPart(part)
default:
return false
}
}
func mergeOutputPartGroup(group []MessageOutputPart) (MessageOutputPart, error) {
if len(group) == 0 {
return MessageOutputPart{}, nil
}
if len(group) == 1 {
return group[0], nil
}
first := group[0]
switch first.Type {
case ChatMessagePartTypeText:
return mergeTextParts(group)
case ChatMessagePartTypeReasoning:
return mergeReasoningParts(group)
case ChatMessagePartTypeAudioURL:
if isBase64MessageOutputAudioPart(first) {
return mergeAudioParts(group)
}
}
return first, nil
}
func mergeTextParts(group []MessageOutputPart) (MessageOutputPart, error) {
var sb strings.Builder
extraList := make([]map[string]any, 0, len(group))
for _, part := range group {
sb.WriteString(part.Text)
if len(part.Extra) > 0 {
extraList = append(extraList, part.Extra)
}
}
var mergedExtra map[string]any
if len(extraList) > 0 {
var err error
mergedExtra, err = concatExtra(extraList)
if err != nil {
return MessageOutputPart{}, fmt.Errorf("failed to concat text part extra: %w", err)
}
}
return MessageOutputPart{
Type: ChatMessagePartTypeText,
Text: sb.String(),
Extra: mergedExtra,
StreamingMeta: group[0].StreamingMeta,
}, nil
}
func mergeReasoningParts(group []MessageOutputPart) (MessageOutputPart, error) {
var textBuilder strings.Builder
var signature string
extraList := make([]map[string]any, 0, len(group))
for _, part := range group {
if part.Reasoning != nil {
textBuilder.WriteString(part.Reasoning.Text)
if part.Reasoning.Signature != "" {
signature = part.Reasoning.Signature
}
}
if len(part.Extra) > 0 {
extraList = append(extraList, part.Extra)
}
}
var mergedExtra map[string]any
if len(extraList) > 0 {
var err error
mergedExtra, err = concatExtra(extraList)
if err != nil {
return MessageOutputPart{}, fmt.Errorf("failed to concat reasoning part extra: %w", err)
}
}
return MessageOutputPart{
Type: ChatMessagePartTypeReasoning,
Reasoning: &MessageOutputReasoning{
Text: textBuilder.String(),
Signature: signature,
},
Extra: mergedExtra,
StreamingMeta: group[0].StreamingMeta,
}, nil
}
func mergeAudioParts(group []MessageOutputPart) (MessageOutputPart, error) {
var b64Builder strings.Builder
var mimeType string
audioExtraList := make([]map[string]any, 0, len(group))
partExtraList := make([]map[string]any, 0, len(group))
for _, part := range group {
audioPart := part.Audio
if audioPart.Base64Data != nil {
b64Builder.WriteString(*audioPart.Base64Data)
}
if mimeType == "" {
mimeType = audioPart.MIMEType
}
if len(audioPart.Extra) > 0 {
audioExtraList = append(audioExtraList, audioPart.Extra)
}
if len(part.Extra) > 0 {
partExtraList = append(partExtraList, part.Extra)
}
}
var mergedAudioExtra map[string]any
var err error
if len(audioExtraList) > 0 {
mergedAudioExtra, err = concatExtra(audioExtraList)
if err != nil {
return MessageOutputPart{}, fmt.Errorf("failed to concat audio extra: %w", err)
}
}
var mergedPartExtra map[string]any
if len(partExtraList) > 0 {
mergedPartExtra, err = concatExtra(partExtraList)
if err != nil {
return MessageOutputPart{}, fmt.Errorf("failed to concat audio part extra: %w", err)
}
}
mergedB64 := b64Builder.String()
return MessageOutputPart{
Type: ChatMessagePartTypeAudioURL,
Audio: &MessageOutputAudio{
MessagePartCommon: MessagePartCommon{
Base64Data: &mergedB64,
MIMEType: mimeType,
Extra: mergedAudioExtra,
},
},
Extra: mergedPartExtra,
StreamingMeta: group[0].StreamingMeta,
}, nil
}
func isBase64MessageOutputAudioPart(part MessageOutputPart) bool {
return part.Type == ChatMessagePartTypeAudioURL &&
part.Audio != nil &&
part.Audio.Base64Data != nil &&
part.Audio.URL == nil
}
func concatUserMultiContent(parts []MessageInputPart) ([]MessageInputPart, error) {
if len(parts) == 0 {
return parts, nil
}
merged := make([]MessageInputPart, 0, len(parts))
i := 0
for i < len(parts) {
currentPart := parts[i]
if currentPart.Type == ChatMessagePartTypeText {
end := i + 1
for end < len(parts) && parts[end].Type == ChatMessagePartTypeText {
end++
}
if end == i+1 {
merged = append(merged, currentPart)
} else {
var sb strings.Builder
for k := i; k < end; k++ {
sb.WriteString(parts[k].Text)
}
mergedPart := MessageInputPart{
Type: ChatMessagePartTypeText,
Text: sb.String(),
}
merged = append(merged, mergedPart)
}
i = end
} else {
merged = append(merged, currentPart)
i++
}
}
return merged, nil
}
func concatExtra(extraList []map[string]any) (map[string]any, error) {
if len(extraList) == 1 {
return generic.CopyMap(extraList[0]), nil
}
return internal.ConcatItems(extraList)
}
// ConcatMessages concat messages with the same role and name.
// It will concat tool calls with the same index.
// It will return an error if the messages have different roles or names.
// It's useful for concatenating messages from a stream.
// e.g.
//
// msgs := []*Message{}
// for {
// msg, err := stream.Recv()
// if errors.Is(err, io.EOF) {
// break
// }
// if err != nil {...}
// msgs = append(msgs, msg)
// }
//
// concatedMsg, err := ConcatMessages(msgs) // concatedMsg.Content will be full content of all messages
func ConcatMessages(msgs []*Message) (*Message, error) {
var (
contents []string
contentLen int
reasoningContents []string
reasoningContentLen int
toolCalls []ToolCall
multiContentParts []ChatMessagePart
assistantGenMultiContentParts []MessageOutputPart
userInputMultiContentParts []MessageInputPart
ret = Message{}
extraList = make([]map[string]any, 0, len(msgs))
)
for idx, msg := range msgs {
if msg == nil {
return nil, fmt.Errorf("unexpected nil chunk in message stream, index: %d", idx)
}
if msg.Role != "" {
if ret.Role == "" {
ret.Role = msg.Role
} else if ret.Role != msg.Role {
return nil, fmt.Errorf("cannot concat messages with "+
"different roles: '%s' '%s'", ret.Role, msg.Role)
}
}
if msg.Name != "" {
if ret.Name == "" {
ret.Name = msg.Name
} else if ret.Name != msg.Name {
return nil, fmt.Errorf("cannot concat messages with"+
" different names: '%s' '%s'", ret.Name, msg.Name)
}
}
if msg.ToolCallID != "" {
if ret.ToolCallID == "" {
ret.ToolCallID = msg.ToolCallID
} else if ret.ToolCallID != msg.ToolCallID {
return nil, fmt.Errorf("cannot concat messages with"+
" different toolCallIDs: '%s' '%s'", ret.ToolCallID, msg.ToolCallID)
}
}
if msg.ToolName != "" {
if ret.ToolName == "" {
ret.ToolName = msg.ToolName
} else if ret.ToolName != msg.ToolName {
return nil, fmt.Errorf("cannot concat messages with"+
" different toolNames: '%s' '%s'", ret.ToolCallID, msg.ToolCallID)
}
}
if msg.Content != "" {
contents = append(contents, msg.Content)
contentLen += len(msg.Content)
}
if msg.ReasoningContent != "" {
reasoningContents = append(reasoningContents, msg.ReasoningContent)
reasoningContentLen += len(msg.ReasoningContent)
}
if len(msg.ToolCalls) > 0 {
toolCalls = append(toolCalls, msg.ToolCalls...)
}
if len(msg.Extra) > 0 {
extraList = append(extraList, msg.Extra)
}
// The 'MultiContent' field is deprecated but is kept for backward compatibility.
if len(msg.MultiContent) > 0 {
multiContentParts = append(multiContentParts, msg.MultiContent...)
}
if len(msg.AssistantGenMultiContent) > 0 {
assistantGenMultiContentParts = append(assistantGenMultiContentParts, msg.AssistantGenMultiContent...)
}
if len(msg.UserInputMultiContent) > 0 {
userInputMultiContentParts = append(userInputMultiContentParts, msg.UserInputMultiContent...)
}
if msg.ResponseMeta != nil && ret.ResponseMeta == nil {
ret.ResponseMeta = &ResponseMeta{}
}
if msg.ResponseMeta != nil && ret.ResponseMeta != nil {
// keep the last FinishReason with a valid value.
if msg.ResponseMeta.FinishReason != "" {
ret.ResponseMeta.FinishReason = msg.ResponseMeta.FinishReason
}
if msg.ResponseMeta.Usage != nil {
if ret.ResponseMeta.Usage == nil {
ret.ResponseMeta.Usage = &TokenUsage{}
}
if msg.ResponseMeta.Usage.PromptTokens > ret.ResponseMeta.Usage.PromptTokens {
ret.ResponseMeta.Usage.PromptTokens = msg.ResponseMeta.Usage.PromptTokens
}
if msg.ResponseMeta.Usage.CompletionTokens > ret.ResponseMeta.Usage.CompletionTokens {
ret.ResponseMeta.Usage.CompletionTokens = msg.ResponseMeta.Usage.CompletionTokens
}
if msg.ResponseMeta.Usage.TotalTokens > ret.ResponseMeta.Usage.TotalTokens {
ret.ResponseMeta.Usage.TotalTokens = msg.ResponseMeta.Usage.TotalTokens
}
if msg.ResponseMeta.Usage.PromptTokenDetails.CachedTokens > ret.ResponseMeta.Usage.PromptTokenDetails.CachedTokens {
ret.ResponseMeta.Usage.PromptTokenDetails.CachedTokens = msg.ResponseMeta.Usage.PromptTokenDetails.CachedTokens
}
if msg.ResponseMeta.Usage.CompletionTokensDetails.ReasoningTokens > ret.ResponseMeta.Usage.CompletionTokensDetails.ReasoningTokens {
ret.ResponseMeta.Usage.CompletionTokensDetails.ReasoningTokens = msg.ResponseMeta.Usage.CompletionTokensDetails.ReasoningTokens
}
}
if msg.ResponseMeta.LogProbs != nil {
if ret.ResponseMeta.LogProbs == nil {
ret.ResponseMeta.LogProbs = &LogProbs{}
}
ret.ResponseMeta.LogProbs.Content = append(ret.ResponseMeta.LogProbs.Content, msg.ResponseMeta.LogProbs.Content...)
}
}
}
if len(contents) > 0 {
var sb strings.Builder
sb.Grow(contentLen)
for _, content := range contents {
_, err := sb.WriteString(content)
if err != nil {
return nil, err
}
}
ret.Content = sb.String()
}
if len(reasoningContents) > 0 {
var sb strings.Builder
sb.Grow(reasoningContentLen)
for _, rc := range reasoningContents {
_, err := sb.WriteString(rc)
if err != nil {
return nil, err
}
}
ret.ReasoningContent = sb.String()
}
if len(toolCalls) > 0 {
merged, err := concatToolCalls(toolCalls)
if err != nil {
return nil, err
}
ret.ToolCalls = merged
}
if len(extraList) > 0 {
extra, err := concatExtra(extraList)
if err != nil {
return nil, fmt.Errorf("failed to concat message's extra: %w", err)
}
if len(extra) > 0 {
ret.Extra = extra
}
}
if len(multiContentParts) > 0 {
ret.MultiContent = multiContentParts
}
if len(assistantGenMultiContentParts) > 0 {
merged, err := concatAssistantMultiContent(assistantGenMultiContentParts)
if err != nil {
return nil, fmt.Errorf("failed to concat message's assistant multicontent: %w", err)
}
ret.AssistantGenMultiContent = merged
}
if len(userInputMultiContentParts) > 0 {
merged, err := concatUserMultiContent(userInputMultiContentParts)
if err != nil {
return nil, fmt.Errorf("failed to concat message's user multicontent: %w", err)
}
ret.UserInputMultiContent = merged
}
return &ret, nil
}
// ConcatMessageStream drains a stream of messages and returns a single
// concatenated message representing the merged content.
func ConcatMessageStream(s *StreamReader[*Message]) (*Message, error) {
defer s.Close()
var msgs []*Message
for {
msg, err := s.Recv()
if err != nil {
if err == io.EOF {
break
}
return nil, err
}
msgs = append(msgs, msg)
}
return ConcatMessages(msgs)
}
// custom jinja env
var jinjaEnvOnce sync.Once
var jinjaEnv *gonja.Environment
var envInitErr error
const (
jinjaInclude = "include"
jinjaExtends = "extends"
jinjaImport = "import"
jinjaFrom = "from"
jinjaFile = "file"
jinjaFileSet = "fileset"
)
func getJinjaEnv() (*gonja.Environment, error) {
jinjaEnvOnce.Do(func() {
jinjaEnv = gonja.NewEnvironment(config.DefaultConfig, gonja.DefaultLoader)
formatInitError := "init jinja env fail: %w"
var err error
if jinjaEnv.Statements.Exists(jinjaInclude) {
err = jinjaEnv.Statements.Replace(jinjaInclude, func(parser *parser.Parser, args *parser.Parser) (nodes.Statement, error) {
return nil, fmt.Errorf("keyword[include] has been disabled")
})
if err != nil {
envInitErr = fmt.Errorf(formatInitError, err)
return
}
}
if jinjaEnv.Statements.Exists(jinjaExtends) {
err = jinjaEnv.Statements.Replace(jinjaExtends, func(parser *parser.Parser, args *parser.Parser) (nodes.Statement, error) {
return nil, fmt.Errorf("keyword[extends] has been disabled")
})
if err != nil {
envInitErr = fmt.Errorf(formatInitError, err)
return
}
}
if jinjaEnv.Statements.Exists(jinjaFrom) {
err = jinjaEnv.Statements.Replace(jinjaFrom, func(parser *parser.Parser, args *parser.Parser) (nodes.Statement, error) {
return nil, fmt.Errorf("keyword[from] has been disabled")
})
if err != nil {
envInitErr = fmt.Errorf(formatInitError, err)
return
}
}
if jinjaEnv.Statements.Exists(jinjaImport) {
err = jinjaEnv.Statements.Replace(jinjaImport, func(parser *parser.Parser, args *parser.Parser) (nodes.Statement, error) {
return nil, fmt.Errorf("keyword[import] has been disabled")
})
if err != nil {
envInitErr = fmt.Errorf(formatInitError, err)
return
}
}
if jinjaEnv.Filters.Exists(jinjaFile) {
err = jinjaEnv.Filters.Replace(jinjaFile, func(e *exec.Evaluator, in *exec.Value, params *exec.VarArgs) *exec.Value {
return exec.AsValue(fmt.Errorf("keyword[file] has been disabled"))
})
if err != nil {
envInitErr = fmt.Errorf(formatInitError, err)
return
}
}
if jinjaEnv.Filters.Exists(jinjaFileSet) {
err = jinjaEnv.Filters.Replace(jinjaFileSet, func(e *exec.Evaluator, in *exec.Value, params *exec.VarArgs) *exec.Value {
return exec.AsValue(fmt.Errorf("keyword[fileset] has been disabled"))
})
if err != nil {
envInitErr = fmt.Errorf(formatInitError, err)
return
}
}
})
return jinjaEnv, envInitErr
}