Eino / LLM

Eino Learning Notes-1-ChatModel

ShemolApr 11, 2025

EinoLLM

ChatModel is Eino's abstraction over conversational LLMs. It gives a unified interface to talk to different model services (OpenAI, Ollama, and so on).

This component matters in these scenes:

Natural-language conversation
Text generation and completion
Generating arguments for tool calls
Multimodal interaction (text, images, audio, etc.)

Component definition

Interface definition

Code location: eino/components/model/interface.go

type ChatModel interface {
    Generate(ctx context.Context, input []*schema.Message, opts ...Option) (*schema.Message, error)
    Stream(ctx context.Context, input []*schema.Message, opts ...Option) (*schema.StreamReader[*schema.Message], error)
    BindTools(tools []*schema.ToolInfo) error
}

Generate method

What it does: generate a full model response
Args:
- ctx: context object, for request-level info, and also for passing the Callback Manager
- input: list of input messages
- opts: optional params to configure model behavior
Returns:
- *schema.Message: the response message the model generated
- error: error info from generation

Stream method

What it does: generate the model response as a stream
Args: same as Generate
Returns:
- *schema.StreamReader[*schema.Message]: streaming reader for the model response
- error: error info from generation

BindTools method

What it does: bind available tools to the model
Args:
- tools: list of tool info
Returns:
- error: error info from binding

Core role: this interface is the core abstraction for chat models. It supports two call modes:

Generate: sync, full response (good for normal chat)
Stream: streaming response (good for long text / realtime)

Architecture notes

type ChatModel interface {
    // 同步生成（典型AI对话模式）
    Generate(ctx context.Context, input []*schema.Message, opts ...Option) (*schema.Message, error)

    // 流式处理（适合逐段输出场景）
    Stream(ctx context.Context, input []*schema.Message, opts ...Option) (
        *schema.StreamReader[*schema.Message], error)

    // 工具绑定机制（支持功能扩展）
    BindTools(tools []*schema.ToolInfo) error
}

Key design points:

Multi-model: the interface abstracts different AI engines (OpenAI/MAAS)
Context-aware: context.Context for timeouts, tracing, etc.
Extensible options: ...Option lets each impl add config
Hot-bind tools: BindTools adds capabilities at runtime (likely Function Calling and similar)

Engineering practice:

A //go:generate directive auto-generates a ChatModelMock, which says:

Interface-first design
Solid unit-test support
Dependency injection (easier to test in different environments)

Things to watch:

Concurrency: comments explicitly say BindTools and Generate are not atomic, so you need sync
Message protocol: depends on the format defined by <schema.Message> (need to look at the actual protocol)
Stream lifecycle: StreamReader needs a Close so resources get released

Message struct

Code location: eino/schema/message.go

type Message struct {   
    // Role 表示消息的角色（system/user/assistant/tool）
    Role RoleType
    // Content 是消息的文本内容
    Content string
    // MultiContent 是多模态内容，支持文本、图片、音频等
    MultiContent []ChatMessagePart
    // Name 是消息的发送者名称
    Name string
    // ToolCalls 是 assistant 消息中的工具调用信息
    ToolCalls []ToolCall
    // ToolCallID 是 tool 消息的工具调用 ID
    ToolCallID string
    // ResponseMeta 包含响应的元信息
    ResponseMeta *ResponseMeta
    // Extra 用于存储额外信息
    Extra map[string]any
}

The Message struct is the basic structure for talking to the model. It supports:

Multiple roles: system, user, assistant (ai), tool
Multimodal content: text, images, audio, video, files
Tool calls: the model can call external tools and functions
Metadata: finish reason, token usage stats, etc.

Shared Option

The Model component provides a set of shared Options to configure behavior:

Code location: eino/components/model/option.go

type Options struct {
    // Temperature 控制输出的随机性
    Temperature *float32
    // MaxTokens 控制生成的最大 token 数量
    MaxTokens *int
    // Model 指定使用的模型名称
    Model *string
    // TopP 控制输出的多样性
    TopP *float32
    // Stop 指定停止生成的条件
    Stop []string
}

You can set Options like this:

// 设置温度
WithTemperature(temperature float32) Option

// 设置最大 token 数
WithMaxTokens(maxTokens int) Option

// 设置模型名称
WithModel(name string) Option

// 设置 top_p 值
WithTopP(topP float32) Option

// 设置停止词
WithStop(stop []string) Option

How to use it

Standalone

import (
    "context"
    "fmt"
    "io"

    "github.com/cloudwego/eino-ext/components/model/openai"
    "github.com/cloudwego/eino/components/model"
    "github.com/cloudwego/eino/schema"
)

// 初始化模型 (以openai为例)
cm, err := openai.NewChatModel(ctx, &openai.ChatModelConfig{
    // 配置参数
})

// 准备输入消息
messages := []*schema.Message{
    {
       Role:    schema.System,
       Content: "你是一个有帮助的助手。",
    },
    {
       Role:    schema.User,
       Content: "你好！",
    },
}

// 生成响应
response, err := cm.Generate(ctx, messages, model.WithTemperature(0.8))

// 响应处理
fmt.Print(response.Content)

// 流式生成
streamResult, err := cm.Stream(ctx, messages)

defer streamResult.Close()

for {
    chunk, err := streamResult.Recv()
    if err == io.EOF {
       break
    }
    if err != nil {
       // 错误处理
    }
    // 响应片段处理
    fmt.Print(chunk.Content)
}

In orchestration

import (
    "github.com/cloudwego/eino/schema"
    "github.com/cloudwego/eino/compose"
)

/*** 初始化ChatModel
* cm, err := xxx
*/

// 在 Chain 中使用
c := compose.NewChain[[]*schema.Message, *schema.Message]()
c.AppendChatModel(cm)


// 在 Graph 中使用
g := compose.NewGraph[[]*schema.Message, *schema.Message]()
g.AddChatModelNode("model_node", cm)

Using Option and Callback

Option example

import "github.com/cloudwego/eino/components/model"

// 使用 Option
response, err := cm.Generate(ctx, messages,
    model.WithTemperature(0.7),
    model.WithMaxTokens(2000),
    model.WithModel("gpt-4"),
)

Callback example

import (
    "context"
    "fmt"

    "github.com/cloudwego/eino/callbacks"
    "github.com/cloudwego/eino/components/model"
    "github.com/cloudwego/eino/compose"
    "github.com/cloudwego/eino/schema"
    callbacksHelper "github.com/cloudwego/eino/utils/callbacks"
)

// 创建 callback handler
handler := &callbacksHelper.ModelCallbackHandler{
    OnStart: func(ctx context.Context, info *callbacks.RunInfo, input *model.CallbackInput) context.Context {
       fmt.Printf("开始生成，输入消息数量: %d\n", len(input.Messages))
       return ctx
    },
    OnEnd: func(ctx context.Context, info *callbacks.RunInfo, output *model.CallbackOutput) context.Context {
       fmt.Printf("生成完成，Token 使用情况: %+v\n", output.TokenUsage)
       return ctx
    },
    OnEndWithStreamOutput: func(ctx context.Context, info *callbacks.RunInfo, output *schema.StreamReader[*model.CallbackOutput]) context.Context {
       fmt.Println("开始接收流式输出")
       defer output.Close()
       return ctx
    },
}

// 使用 callback handler
helper := callbacksHelper.NewHandlerHelper().
    ChatModel(handler).
    Handler()

/*** compose a chain
* chain := NewChain
* chain.appendxxx().
*       appendxxx().
*       ...
*/

// 在运行时使用
runnable, err := chain.Compile()
if err != nil {
    return err
}
result, err := runnable.Invoke(ctx, messages, compose.WithCallbacks(helper))

Existing implementations

OpenAI ChatModel: GPT-series models from OpenAI ChatModel - OpenAI
Ollama ChatModel: local models via Ollama ChatModel - Ollama
ARK ChatModel: models on the ARK platform ChatModel - ARK

Implementing your own

When you implement a custom ChatModel, watch these:

Implement the shared options
Implement the callback mechanism
On streaming, remember to close the writer when output is done

Option mechanism

If a custom ChatModel needs Options beyond the shared ones, you can build custom Options with the component helper functions, for example:

import (
    "time"

    "github.com/cloudwego/eino/components/model"
)

// 定义 Option 结构体
type MyChatModelOptions struct {
    Options    *model.Options
    RetryCount int
    Timeout    time.Duration
}

// 定义 Option 函数
func WithRetryCount(count int) model.Option {
    return model.WrapImplSpecificOptFn(func(o *MyChatModelOptions) {
       o.RetryCount = count
    })
}

func WithTimeout(timeout time.Duration) model.Option {
    return model.WrapImplSpecificOptFn(func(o *MyChatModelOptions) {
       o.Timeout = timeout
    })
}

Handling Callback

A ChatModel impl needs to fire callbacks at the right times. These structs are defined by the ChatModel component:

import (
    "github.com/cloudwego/eino/schema"
)

// 定义回调输入输出
type CallbackInput struct {
    Messages    []*schema.Message
    Model       string
    Temperature *float32
    MaxTokens   *int
    Extra       map[string]any
}

type CallbackOutput struct {
    Message    *schema.Message
    TokenUsage *schema.TokenUsage
    Extra      map[string]any
}

Full implementation example

import (
    "context"
    "errors"
    "net/http"
    "time"

    "github.com/cloudwego/eino/callbacks"
    "github.com/cloudwego/eino/components/model"
    "github.com/cloudwego/eino/schema"
)

type MyChatModel struct {
    client     *http.Client
    apiKey     string
    baseURL    string
    model      string
    timeout    time.Duration
    retryCount int
}

type MyChatModelConfig struct {
    APIKey string
}

func NewMyChatModel(config *MyChatModelConfig) (*MyChatModel, error) {
    if config.APIKey == "" {
       return nil, errors.New("api key is required")
    }

    return &MyChatModel{
       client: &http.Client{},
       apiKey: config.APIKey,
    }, nil
}

func (m *MyChatModel) Generate(ctx context.Context, messages []*schema.Message, opts ...model.Option) (*schema.Message, error) {
    // 1. 处理选项
    options := &MyChatModelOptions{
       Options: &model.Options{
          Model: &m.model,
       },
       RetryCount: m.retryCount,
       Timeout:    m.timeout,
    }
    options.Options = model.GetCommonOptions(options.Options, opts...)
    options = model.GetImplSpecificOptions(options, opts...)

    // 2. 开始生成前的回调
    ctx = callbacks.OnStart(ctx, &model.CallbackInput{
       Messages: messages,
       Config: &model.Config{
          Model: *options.Options.Model,
       },
    })

    // 3. 执行生成逻辑
    response, err := m.doGenerate(ctx, messages, options)

    // 4. 处理错误和完成回调
    if err != nil {
       ctx = callbacks.OnError(ctx, err)
       return nil, err
    }

    ctx = callbacks.OnEnd(ctx, &model.CallbackOutput{
       Message: response,
    })

    return response, nil
}

func (m *MyChatModel) Stream(ctx context.Context, messages []*schema.Message, opts ...model.Option) (*schema.StreamReader[*schema.Message], error) {
    // 1. 处理选项
    options := &MyChatModelOptions{
       Options: &model.Options{
          Model: &m.model,
       },
       RetryCount: m.retryCount,
       Timeout:    m.timeout,
    }
    options.Options = model.GetCommonOptions(options.Options, opts...)
    options = model.GetImplSpecificOptions(options, opts...)

    // 2. 开始流式生成前的回调
    ctx = callbacks.OnStart(ctx, &model.CallbackInput{
       Messages: messages,
       Config: &model.Config{
          Model: *options.Options.Model,
       },
    })

    // 3. 创建流式响应
    // Pipe产生一个StreamReader和一个StreamWrite，向StreamWrite中写入可以从StreamReader中读到，二者并发安全。
    // 实现中异步向StreamWrite中写入生成内容，返回StreamReader作为返回值
    // ***StreamReader是一个数据流，仅可读一次，组件自行实现Callback时，既需要通过OnEndWithCallbackOutput向callback传递数据流，也需要向返回一个数据流，需要对数据流进行一次拷贝
    // 考虑到此种情形总是需要拷贝数据流，OnEndWithCallbackOutput函数会在内部拷贝并返回一个未被读取的流
    // 以下代码演示了一种流处理方式，处理方式不唯一
    sr, sw := schema.Pipe[*model.CallbackOutput](1)

    // 4. 启动异步生成
    go func() {
       defer sw.Close()

       // 流式写入
       m.doStream(ctx, messages, options, sw)
    }()

    // 5. 完成回调
    _, nsr := callbacks.OnEndWithStreamOutput(ctx, sr)

    return schema.StreamReaderWithConvert(nsr, func(t *model.CallbackOutput) (*schema.Message, error) {
       return t.Message, nil
    }), nil
}

func (m *MyChatModel) BindTools(tools []*schema.ToolInfo) error {
    // 实现工具绑定逻辑
    return nil
}

func (m *MyChatModel) doGenerate(ctx context.Context, messages []*schema.Message, opts *MyChatModelOptions) (*schema.Message, error) {
    // 实现生成逻辑
    return nil, nil
}

func (m *MyChatModel) doStream(ctx context.Context, messages []*schema.Message, opts *MyChatModelOptions, sr *schema.StreamWriter[*model.CallbackOutput]) {
    // 流式生成文本写入sr中
    return
}