Defining AI workflows

AI workflows typically require more than just a model call. They need pre- and post-processing steps like retrieving context, managing session history, reformatting inputs, validating outputs, or combining multiple model responses.

A flow is a special Genkit function that wraps your AI logic to provide:

• Type-safe inputs and outputs: Define schemas using Go structs for static and runtime validation
• Streaming support: Stream partial responses or custom data
• Developer UI integration: Test and debug flows with visual traces
• Easy deployment: Deploy as HTTP endpoints to any platform

Flows are lightweight. They’re written like regular functions with minimal abstraction.

Defining and calling flows

In its simplest form, a flow just wraps a function. The following example wraps a function that makes a model generation request:

package main

import (
    "context"
    "fmt"

    "github.com/firebase/genkit/go/ai"
    "github.com/firebase/genkit/go/genkit"
    "github.com/firebase/genkit/go/plugins/googlegenai"
)

func main() {
    ctx := context.Background()

    g := genkit.Init(ctx,
        genkit.WithPlugins(&googlegenai.GoogleAI{}),
    )

    type MenuSuggestionInput struct {
        Theme string `json:"theme"`
    }

    type MenuSuggestionOutput struct {
        MenuItem string `json:"menuItem"`
    }

    genkit.DefineFlow(g, "menuSuggestionFlow",
        func(ctx context.Context, input MenuSuggestionInput) (MenuSuggestionOutput, error) {
            resp, err := genkit.Generate(ctx, g,
                ai.WithPrompt("Invent a menu item for a %s themed restaurant.", input.Theme),
            )
            if err != nil {
                return MenuSuggestionOutput{}, fmt.Errorf("could not invent a menu item: %w", err)
            }

            return MenuSuggestionOutput{MenuItem: resp.Text()}, nil
        })
}

Just by wrapping your generate calls like this, you add some functionality: doing so lets you run the flow from the Genkit CLI and from the developer UI, and is a requirement for several of Genkit’s features, including deployment and observability (later sections discuss these topics).

Input and output schemas

One of the most important advantages Genkit flows have over directly calling a model API is type safety of both inputs and outputs. When defining flows, you can define schemas for them.

You can define schemas using Go structs with JSON tags. While you can use primitive types directly as input and output parameters, it’s considered best practice to use struct-based schemas for these reasons:

• Better developer experience: Struct-based schemas provide a better experience in the Developer UI by giving you labeled input fields.
• Future-proof API design: Struct-based schemas allow for easy extensibility in the future. You can add new fields to your input or output schemas without breaking existing clients, which is a core principle of robust API design.

All examples in this documentation use struct-based schemas to follow these best practices.

Here’s a refinement of the last example, which defines a flow that takes a string as input and outputs an object:

type MenuSuggestionInput struct {
    Theme string `json:"theme"`
}

type MenuItem struct {
    Name        string `json:"name"`
    Description string `json:"description"`
}

menuSuggestionFlow := genkit.DefineFlow(g, "menuSuggestionFlow",
    func(ctx context.Context, input MenuSuggestionInput) (MenuItem, error) {
        item, resp, err := genkit.GenerateData[MenuItem](ctx, g,
            ai.WithPrompt("Invent a menu item for a %s themed restaurant.", input.Theme),
        )
        if err != nil {
            return MenuItem{}, fmt.Errorf("could not invent a menu item: %w", err)
        }
        // GenerateData answers with a nil value and no error when the
        // response carried no text to parse, as an interrupted one does.
        if item == nil {
            return MenuItem{}, status.Errorf(status.ErrInternal,
                "the model returned no menu item (%s)", resp.FinishReason)
        }

        return *item, nil
    })

genkit.GenerateData[MenuItem] returns a *MenuItem, so a caller has three things to check in order: the error, then whether the value is nil, and only then the value itself. A nil value with no error means the response carried something other than text, such as a tool request or an interrupt, so inspect resp.FinishReason, resp.Interrupts(), or resp.ToolRequests() to decide what to do. The basic-structured sample works through typed output, including the nested and streaming cases.

Note that the schema of a flow does not necessarily have to line up with the schema of the model generation calls within the flow (in fact, a flow might not even contain model calls). Here’s a variation of the example that uses the structured output to format a simple string, which the flow returns.

Note how we pass MenuItem as a type parameter; this is the equivalent of passing the WithOutputType() option and getting a value of that type in response.

type MenuSuggestionInput struct {
    Theme string `json:"theme"`
}

type MenuItem struct {
    Name        string `json:"name"`
    Description string `json:"description"`
}

type FormattedMenuOutput struct {
    FormattedMenuItem string `json:"formattedMenuItem"`
}

menuSuggestionMarkdownFlow := genkit.DefineFlow(g, "menuSuggestionMarkdownFlow",
    func(ctx context.Context, input MenuSuggestionInput) (FormattedMenuOutput, error) {
        item, resp, err := genkit.GenerateData[MenuItem](ctx, g,
            ai.WithPrompt("Invent a menu item for a %s themed restaurant.", input.Theme),
        )
        if err != nil {
            return FormattedMenuOutput{}, fmt.Errorf("could not invent a menu item: %w", err)
        }
        if item == nil {
            return FormattedMenuOutput{}, status.Errorf(status.ErrInternal,
                "the model returned no menu item (%s)", resp.FinishReason)
        }

        return FormattedMenuOutput{
            FormattedMenuItem: fmt.Sprintf("**%s**: %s", item.Name, item.Description),
        }, nil
    })

Calling flows

Once you’ve defined a flow, you can call it from your code:

output, err := menuSuggestionFlow.Run(context.Background(), MenuSuggestionInput{Theme: "bistro"})

The argument to the flow must conform to the input schema.

If you defined an output schema, the flow response will conform to it. For example, if you set the output schema to MenuItemSchema, the flow output will contain its properties:

item, err := menuSuggestionFlow.Run(context.Background(), MenuSuggestionInput{Theme: "bistro"})
if err != nil {
    log.Fatal(err)
}

log.Println(item.Name)
log.Println(item.Description)

Streaming flows

Flows support streaming using an interface similar to the model generation streaming interface. Streaming is useful when your flow generates a large amount of output, because you can present the output to the user as it’s being generated, which improves the perceived responsiveness of your app. As a familiar example, chat-based LLM interfaces often stream their responses to the user as they are generated.

Durable streaming

For flows that run for a long time or where network reliability is a concern, you can use durable streaming. This allows the client to reconnect to a stream and replay the content. Note that durable streaming for Go is currently experimental.

Here’s an example of a flow that supports streaming:

A streaming flow is defined with genkit.DefineStreamingFlow, whose function takes a sendChunk callback alongside the input. What that function looks like depends on what the flow does with the model’s output. The basic sample shows the forwarding shape, and the basic-structured sample has both shapes side by side.

Forwarding the model’s chunks

When the flow passes the model’s output straight through, ai.WithStreaming is the whole job: hand it the flow’s own sendChunk and the model’s chunks reach the caller untouched, while genkit.Generate still returns the finished response.

type MenuSuggestionInput struct {
    Theme string `json:"theme"`
}

menuSuggestionFlow := genkit.DefineStreamingFlow(g, "menuSuggestionFlow",
    func(ctx context.Context, input MenuSuggestionInput, sendChunk ai.ModelStreamCallback) (string, error) {
        resp, err := genkit.Generate(ctx, g,
            ai.WithPrompt("Invent a menu item for a %s themed restaurant.", input.Theme),
            ai.WithStreaming(sendChunk),
        )
        if err != nil {
            return "", fmt.Errorf("could not invent a menu item: %w", err)
        }

        return resp.Text(), nil
    })

Declaring the callback as ai.ModelStreamCallback is what lets sendChunk be passed straight through: the flow streams the model’s chunks unchanged, so it streams the model’s chunk type. To stream something else, wrap the callback and declare the type you want:

type Menu struct {
    Theme string     `json:"theme"`
    Items []MenuItem `json:"items"`
}

type MenuItem struct {
    Name        string `json:"name"`
    Description string `json:"description"`
}

menuSuggestionFlow := genkit.DefineStreamingFlow(g, "menuSuggestionFlow",
    func(ctx context.Context, input MenuSuggestionInput, sendChunk core.StreamCallback[string]) (Menu, error) {
        item, resp, err := genkit.GenerateData[MenuItem](ctx, g,
            ai.WithPrompt("Invent a menu item for a %s themed restaurant.", input.Theme),
            ai.WithStreaming(func(ctx context.Context, chunk *ai.ModelResponseChunk) error {
                return sendChunk(ctx, chunk.Text())
            }),
        )
        if err != nil {
            return Menu{}, fmt.Errorf("could not invent a menu item: %w", err)
        }
        if item == nil {
            return Menu{}, status.Errorf(status.ErrInternal,
                "the model returned no menu item (%s)", resp.FinishReason)
        }

        return Menu{
            Theme: input.Theme,
            Items: []MenuItem{*item},
        }, nil
    })

The string in core.StreamCallback[string] is the type the flow streams. It does not have to be the same type as the flow’s complete output (Menu here). chunk.Text() is the text that just arrived rather than the message so far, so forwarding it gives a client something to append.

Acting on the chunks

When the flow has to inspect, filter, or accumulate what arrives, range over genkit.GenerateStream, which hands you each chunk and then the final response in one loop:

menuSuggestionFlow := genkit.DefineStreamingFlow(g, "menuSuggestionFlow",
    func(ctx context.Context, input MenuSuggestionInput, sendChunk core.StreamCallback[string]) (string, error) {
        for result, err := range genkit.GenerateStream(ctx, g,
            ai.WithPrompt("Invent a menu item for a %s themed restaurant.", input.Theme),
        ) {
            if err != nil {
                return "", fmt.Errorf("could not invent a menu item: %w", err)
            }
            if result.Done {
                return result.Response.Text(), nil
            }
            // Forward only the text, dropping the tool traffic a multi-turn
            // run also carries.
            if text := result.Chunk.Text(); text != "" {
                sendChunk(ctx, text)
            }
        }

        return "", status.Errorf(status.ErrInternal, "the stream ended without a final result")
    })

For typed output, genkit.GenerateDataStream[T] does the same with the schema inferred from T:

menuSuggestionFlow := genkit.DefineStreamingFlow(g, "menuSuggestionFlow",
    func(ctx context.Context, input MenuSuggestionInput, sendChunk core.StreamCallback[*MenuItem]) (*MenuItem, error) {
        for result, err := range genkit.GenerateDataStream[*MenuItem](ctx, g,
            ai.WithPrompt("Invent a menu item for a %s themed restaurant.", input.Theme),
        ) {
            if err != nil {
                return nil, fmt.Errorf("could not invent a menu item: %w", err)
            }
            if result.Done {
                // result.Output is strongly typed as *MenuItem.
                return result.Output, nil
            }
            // result.Chunk is the whole MenuItem parsed so far.
            sendChunk(ctx, result.Chunk)
        }

        return nil, status.Errorf(status.ErrInternal, "the stream ended without a final result")
    })

With the default JSON format each chunk replaces the previous one rather than adding to it: the whole value is reparsed from everything received so far, so fields fill in as the model writes them. A chunk that has not parsed into anything yet is skipped when the type parameter is a pointer, so result.Chunk is never nil here. With a value type such as MenuItem those chunks arrive as the zero value instead.

The values a flow streams are often coupled to the values the generation call inside it streams, but they do not have to be. You can call sendChunk as often as is useful, with whatever your flow has to report.

Using channel-based streaming (experimental)

Experimental API

The channel-based streaming API is in preview and may change in any minor release. It lives in github.com/firebase/genkit/go/genkit/exp and requires genkit.WithExperimental() at initialization.

For a more idiomatic Go approach, you can use the experimental channel-based streaming API. Instead of a callback, your flow function receives a channel to which it writes stream chunks:

import (
    "context"
    "fmt"

    "github.com/firebase/genkit/go/ai"
    "github.com/firebase/genkit/go/core/status"
    "github.com/firebase/genkit/go/genkit"
    genkitx "github.com/firebase/genkit/go/genkit/exp"
)

jokeFlow := genkitx.DefineStreamingFlow(g, "jokeFlow",
    func(ctx context.Context, topic string, streamCh chan<- string) (string, error) {
        for result, err := range genkit.GenerateStream(ctx, g,
            ai.WithPrompt("Tell me a joke about %s.", topic),
        ) {
            if err != nil {
                return "", fmt.Errorf("could not generate joke: %w", err)
            }
            if result.Done {
                return result.Response.Text(), nil
            }
            select {
            case streamCh <- result.Chunk.Text():
            case <-ctx.Done():
                return "", ctx.Err()
            }
        }

        return "", status.Errorf(status.ErrInternal, "the stream ended without a final result")
    })

The channel-based API:

• Passes a send-only channel (chan<- string) to your function instead of a callback
• The channel is managed by the framework and closed automatically after your function returns
• Your function should NOT close the channel
• Use a select statement with ctx.Done() to handle cancellation gracefully

The returned flow works identically to callback-based flows and can be run with either Run() or Stream().

Calling streaming flows

Streaming flows can be run like non-streaming flows with menuSuggestionFlow.Run(ctx, MenuSuggestionInput{Theme: "bistro"}) or they can be streamed:

for result, err := range menuSuggestionFlow.Stream(context.Background(), MenuSuggestionInput{Theme: "bistro"}) {
    if err != nil {
        log.Fatalf("Stream error: %v", err)
    }
    if result.Done {
        log.Printf("Menu with %s theme:\n", result.Output.Theme)
        for _, item := range result.Output.Items {
            log.Printf(" - %s: %s", item.Name, item.Description)
        }
    } else {
        log.Println("Stream chunk:", result.Stream)
    }
}

Stream returns an iterator, so the error arrives as the loop’s second value. result.Stream carries a chunk while result.Done is false, and result.Output carries the flow’s complete output once it is true.

Handling errors in flows

A flow returns an ordinary Go error. Classify it once, where its meaning is known, using a sentinel from the core/status package. status.PublicErrorf marks a message as safe to show a client; status.Errorf keeps it internal.

if strings.TrimSpace(input.Dish) == "" {
    return "", status.PublicErrorf(status.ErrInvalidArgument, "dish must not be empty")
}

Add context as the error travels with fmt.Errorf and %w. Wrapping does not reclassify, so the status and the public message chosen at the source survive to the HTTP boundary. %v flattens the error to text and loses both.

recipe, err := lookupRecipe(input.Dish)
if err != nil {
    return "", fmt.Errorf("could not look up the recipe: %w", err)
}

Branch with errors.Is against a sentinel, never on message text:

if errors.Is(err, ErrRecipeNotFound) {
    // Recover: improvise a recipe instead of failing the request.
}

At the HTTP boundary the status picks the response code, and only a message written with status.PublicErrorf reaches the client. Anything else is replaced with a generic string and logged in full server-side. Setting GENKIT_ENV=dev returns the real message instead, so what you see while developing is not the production contract. The status code is the same either way. See Error types for the sentinels, subtypes, and the redaction rules. The basic-errors sample walks one flow that produces classified errors, one that recovers from them, and one that fails unclassified so you can watch the boundary redact it.

Errors on a streaming request

Asking for a stream changes where a failure lands. The server has already answered 200 OK with Content-Type: text/event-stream before the flow runs, so the failure arrives in the body as a final event rather than on the status line:

data: {"error":{"status":"INVALID_ARGUMENT","message":"dish must not be empty"}}

The frame carries exactly two fields, status and message, and is terminated by a blank line like every other event. Chunks sent before the failure are still delivered, so a client must read to the end of the stream rather than key on the HTTP status.

Running flows from the command line

You can run flows from the command line using the Genkit CLI tool:

genkit flow:run menuSuggestionFlow '{"theme": "French"}' -- <command to start your app>

For streaming flows, you can print the streaming output to the console by adding the -s flag:

genkit flow:run menuSuggestionFlow '{"theme": "French"}' -s -- <command to start your app>

Running a flow from the command line is useful for testing a flow, or for running flows that perform tasks needed on an ad hoc basis—for example, to run a flow that ingests a document into your vector database.

Debugging flows

One of the advantages of encapsulating AI logic within a flow is that you can test and debug the flow independently from your app using the Genkit developer UI.

The developer UI relies on the Go app continuing to run, even if the logic has completed. If you are just getting started and Genkit is not part of a broader app, add select {} as the last line of main() to prevent the app from shutting down so that you can inspect it in the UI.

To start the developer UI, run the following command from your project directory:

genkit start -- go run .

From the Run tab of developer UI, you can run any of the flows defined in your project:

After you’ve run a flow, you can inspect a trace of the flow invocation by either clicking View trace or looking at the Inspect tab.

Flow steps

Each of Genkit’s fundamental actions show up as separate steps in the trace viewer:

• genkit.Generate()
• genkit.Embed()
• genkit.Retrieve()

If you want to include code other than the above in your traces, you can do so by wrapping the code in a genkit.RunWithContext() call. You might do this for calls to third-party libraries that are not Genkit-aware, or for any critical section of code.

For example, here’s a flow with two steps: the first step retrieves a menu using some unspecified method, and the second step includes the menu as context for a genkit.Generate() call.

type MenuQuestionInput struct {
    Question string `json:"question"`
}

type MenuQuestionOutput struct {
    Answer string `json:"answer"`
}

menuQuestionFlow := genkit.DefineFlow(g, "menuQuestionFlow",
    func(ctx context.Context, input MenuQuestionInput) (MenuQuestionOutput, error) {
        menu, err := genkit.RunWithContext(ctx, "retrieve-daily-menu",
            func(ctx context.Context) (string, error) {
                // Retrieve today's menu. (This could be a database access or
                // simply fetching the menu from your website.)
                return fetchMenu(ctx)
            })
        if err != nil {
            return MenuQuestionOutput{}, fmt.Errorf("could not retrieve the menu: %w", err)
        }

        resp, err := genkit.Generate(ctx, g,
            ai.WithPrompt(input.Question),
            ai.WithSystem("Help the user answer questions about today's menu."),
            ai.WithTextDocs(menu),
        )
        if err != nil {
            return MenuQuestionOutput{}, fmt.Errorf("could not answer the question: %w", err)
        }

        return MenuQuestionOutput{Answer: resp.Text()}, nil
    })

RunWithContext hands the step its own context, so anything the step starts with it, an HTTP client or a database call, nests under the step in the trace instead of beside it under the flow. Use genkit.Run instead when the step is pure work that traces nothing of its own; it takes a func() (Out, error) with no context.

Because the retrieval step is wrapped in a genkit.RunWithContext() call, it’s included as a step in the trace viewer:

Deploying flows

You can deploy your flows directly as web API endpoints, ready for you to call from your app clients. Deployment is discussed in detail on several other pages, but this section gives brief overviews of your deployment options.

net/http server

To deploy a flow using any Go hosting platform, such as Cloud Run, define your flow using genkit.DefineFlow() and start a net/http server with the provided flow handler using genkit.Handler():

package main

import (
    "context"
    "fmt"
    "log"
    "net/http"

    "github.com/firebase/genkit/go/ai"
    "github.com/firebase/genkit/go/core/status"
    "github.com/firebase/genkit/go/genkit"
    "github.com/firebase/genkit/go/plugins/googlegenai"
    "github.com/firebase/genkit/go/plugins/server"
)

type MenuSuggestionInput struct {
    Theme string `json:"theme"`
}

type MenuItem struct {
    Name        string `json:"name"`
    Description string `json:"description"`
}

func main() {
    ctx := context.Background()

    g := genkit.Init(ctx, genkit.WithPlugins(&googlegenai.GoogleAI{}))

    menuSuggestionFlow := genkit.DefineFlow(g, "menuSuggestionFlow",
        func(ctx context.Context, input MenuSuggestionInput) (MenuItem, error) {
            item, resp, err := genkit.GenerateData[MenuItem](ctx, g,
                ai.WithPrompt("Invent a menu item for a %s themed restaurant.", input.Theme),
            )
            if err != nil {
                return MenuItem{}, fmt.Errorf("could not invent a menu item: %w", err)
            }
            if item == nil {
                return MenuItem{}, status.Errorf(status.ErrInternal,
                    "the model returned no menu item (%s)", resp.FinishReason)
            }
            return *item, nil
        })

    mux := http.NewServeMux()
    mux.HandleFunc("POST /menuSuggestionFlow", genkit.Handler(menuSuggestionFlow))
    log.Fatal(server.Start(ctx, "127.0.0.1:3400", mux))
}

server.Start() is an optional helper function that starts the server and manages its lifecycle, including capturing interrupt signals to ease local development, but you may use your own method.

To serve all the flows defined in your codebase, you can use genkit.ListFlows():

mux := http.NewServeMux()
for _, flow := range genkit.ListFlows(g) {
    mux.HandleFunc("POST /"+flow.Name(), genkit.Handler(flow))
}
log.Fatal(server.Start(ctx, "127.0.0.1:3400", mux))

Other server frameworks

You can also use other server frameworks to deploy your flows. For example, you can use Gin with just a few lines:

router := gin.Default()
for _, flow := range genkit.ListFlows(g) {
    router.POST("/"+flow.Name(), func(c *gin.Context) {
        genkit.Handler(flow)(c.Writer, c.Request)
    })
}
log.Fatal(router.Run(":3400"))

Calling deployed flows

Once your flow is deployed, you can call it with a POST request:

curl -X POST "http://localhost:3400/menuSuggestionFlow" \
  -H "Content-Type: application/json" -d '{"data": {"theme": "banana"}}'

For streaming responses, you can add the Accept: text/event-stream header:

curl -X POST "http://localhost:3400/menuSuggestionFlow" \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{"data": {"theme": "banana"}}'

Learn more about deployment

For detailed deployment instructions and platform-specific guides, see:

• Deploy with Cloud Run