genkit-go-flows

Genkit for Go: Flows, Prompts, and Tools

Genkit (github.com/firebase/genkit/go) is Google/Firebase's application framework for building AI features — flows, prompt management, structured generation, tool calling, RAG, and evals — with first-class Go support alongside JS/TS. It is not the same tool as adk-go-agent-builder's Agent Development Kit: Genkit is the broader app-framework layer (any AI feature, not just autonomous agents), while ADK is specifically for building orchestrated, tool-using agents. A service can reasonably use both — Genkit for its prompt/generation plumbing and observability, ADK for the parts that are genuinely agentic.

Defining a flow

A flow is Genkit's typed, traceable unit of AI-feature logic — register it against a *genkit.Genkit instance rather than calling generation functions ad hoc, so it shows up in traces and can be run standalone via the CLI:

g := genkit.Init(ctx)

flow := genkit.DefineFlow(g, "summarizeReservation", func(ctx context.Context, input ReservationInput) (string, error) {
    resp, err := genkit.Generate(ctx, g,
        ai.WithPrompt("Summarize this reservation in one sentence: %+v", input),
    )
    if err != nil {
        return "", fmt.Errorf("generate summary: %w", err)
    }
    return resp.Text(), nil
})

Pass g (the *Genkit instance) explicitly into every call that needs it — it is Genkit's central action registry (flows, prompts, tools all register against it). Storing it as a package-level global instead of threading it through breaks the documented pattern and makes a service harder to test with an isolated registry per test.

Running with traces (the dev loop that actually works)

genkit start -- go run .              # traced, with Dev UI
genkit flow:run summarizeReservation '{"hotelId":"h1"}' -- go run .   # traced, non-interactive

go run . on its own does not capture Genkit dev traces — only running under genkit start (or the equivalent genkit flow:run for a single, self-terminating invocation) wires up the trace collector. Debugging a flow by running the binary directly is debugging blind: no trace, no Dev UI inspection of the actual prompt/response the model saw.