README

AI Worksflows with LiveBindings and Replicate

Build visual AI workflows in Delphi FMX — wire Replicate.com models together in the LiveBindings Designer, press Run, and watch the chain execute.

-6E4C9E)

Drop model nodes on a form, drag a line from one model's output to the next model's input, and the workflow runs itself: generate an image with flux-schnell, and the upscaler downstream launches automatically the moment the URL lands in its input. Inputs are typed from each model's JSON schema, outputs fan out to grids and images, and a JSON-aware relay picks which field of an output travels on down the chain.

?????????????????????????????????????????????????????????      ????????????????????????????????????????????????????????????      ?????????????????????????????????????????????????????????????????????
???  flux-schnell   ????????????????????????  > output[0] >   ???????????????????????? recraft-crisp-      ???
???  "a red car..." ???      ???  (TReplicateRelay)???      ??? upscale   AutoRun   ???
?????????????????????????????????????????????????????????      ????????????????????????????????????????????????????????????      ?????????????????????????????????????????????????????????????????????
   TReplicateNode          JSON field picker           TReplicateNode
        ??????????????????????????????????????? all three are LiveBindings Designer lines ????????????????????????????????????

Contents

• Why this project
• Components
• Quick start
• Your first workflow
• Samples
• Repository layout
• Requirements
• Notes & caveats
• License

Why this project

• Visual, not glue code. Model→model chains are drawn in the LiveBindings Designer, so the topology of a workflow is visible on the form — no hidden event wiring.
• Schema-driven. Inputs are typed straight from each model's JSON schema; TReplicateInputPanel builds a full input form (switches, sliders, enum-aware combos, file pickers) at runtime with zero configuration.
• Cascading execution. Set AutoRun and a downstream stage launches itself the instant its input arrives — one node fanning into the next into the next.
• Code-first when you want it. Every visual capability has a plain-code path (Chat.PipeTextTo(Gen).PipeImageTo(Up)) for pipelines without bindings.
• Mixed providers. An optional SmartCore chat engine (OpenAI / Claude / Gemini / Ollama) plugs into the same nodes and chains as the Replicate models.

Components

Design-time extras: a model picker (searches Replicate's catalog), Load Model Schema (caches the schema into the form file), Map Schemas..., and an experimental Generate Input Form verb that emits controls + field links onto the form.

Quick start

1. Build & install the packages (order matters)

• Open Packages/LiveReplicate.dproj → Build (runtime).
• Open Packages/dclLiveReplicate.dproj → Build, then Install (design-time).
• Restart the IDE. A Replicate palette page appears.

2. Set your token

Preferred: the REPLICATE_API_TOKEN environment variable — the components and all samples fall back to it.

[!WARNING] Never commit a token. Form files are code, and a pasted token is saved with them. Get a token at replicate.com/account/api-tokens.

3. Run a sample

Start with Samples/08-Nodes for the node workflow, or see the samples index for the full learning path.

Your first workflow, from scratch

1. Drop two TReplicateNodes on an FMX form. Right-click each → Pick Replicate Model... (e.g. black-forest-labs/flux-schnell and an upscaler), then Load Model Schema.
2. Open the LiveBindings Designer and drag NodeGen.OutputImage → NodeUpscale.InputImage.
3. Set NodeUpscale.AutoRun := True.
4. In a button click:
   pascalCopy

   NodeGen.SetInputValue('prompt', Edit1.Text);
   NodeGen.Run;

Stage 2 launches itself when stage 1 completes. Bind Status, Logs, and OutputImage to controls for live progress; use a TReplicateRelay between nodes when you need to pick a specific field out of the output JSON.

Samples

The Samples/ folder is a numbered learning path — each sample is self-contained: open the .dproj, set REPLICATE_API_TOKEN, run.

Start with 08-Nodes, then 09-Schema and 10-Relay — those three are the current architecture. 12-PureBindings is the payoff: a three-stage pipeline with not a single event handler in the unit. Numbering gaps are deliberate — they belong to superseded approaches, preserved under Samples/archive with notes on why each was outgrown.

Repository layout

Source/     Runtime components (one unit per component)
Packages/   LiveReplicate (runtime) + dclLiveReplicate (design-time)
Samples/    Numbered learning path (superseded approaches under Samples/archive)

Requirements

• Delphi 13 with FireMonkey (also builds on Delphi 12.x Athens)
• FireDAC — in-memory tables only, no database
• A Replicate API token
• Internet access at runtime
• For mixed-provider workflows: the SmartCore AI pack (from GetIt) for the TSmartCoreChatEngine and the SmartFlow sample

Notes & caveats

• Developed against Delphi 13 and Delphi 12.x. Designer-integration behaviors are version-sensitive.
• Predictions run on Replicate's infrastructure and cost money — the components guard against accidental design-time launches, and AutoRun only fires on non-empty inputs, but review chains before running them.
• Win32/Win64 project targets are configured; other FMX platforms should work but are untested.

Origins

The initial project was scaffolded in RAD Studio Kai, Embarcadero's AI-assisted development for Delphi, and has been developed further since.

License

Released under the MIT License — © 2026 FMXExpress.