Gemini LLM client for Delphi

1 Comment / By /

I open-sourced my AI client. For the moment it only supports Google Gemini, but it can be easily extended to other LLMs (Claude, OpenAI, etc.) by subclassing TLLMObject.

What it does

This Delphi library lets your application talk to the Gemini API without you having to fight with HTTP, multipart upload sessions, JSON body assembly, or response parsing. It handles all of that and gives you back a clean, structured result.

Library features:

Cross-platform FMX — works on Windows, macOS, Linux, iOS, Android
Tested with the latest version of Delphi
Heavily tested in production

Code features:

Send text prompts and get back structured JSON responses (enforced via a response schema)
Upload local files (images, documents, text) and reference them in the prompt — uses Google’s resumable upload protocol
Send multiple files to the AI in a single request
Multi-turn chat support
Ready-made FMX settings dialog: API key, model, Temperature, TopP, TopK, MaxTokens, and the “thinking” toggle for Gemini 2.5+
Settings and total token usage persisted between sessions (binary serialization to AppDataFolder)
Built-in connection test, with meaningful error reporting (HTTP errors, safety blocks, token limits, etc.)
Every AI response JSON is backed up to disk — handy for debugging and replay
Safety parameters to reject inappropriate input

Class architecture

TAiClient                          Low-level HTTP, file upload, response parsing
  +-- TAiClientEx                  High-level SendAPICall, JSON body, multi-turn
        +-- LLM: TLLMObject        Abstract base for any LLM provider
                  +-- TLLMGemini   Gemini URLs, defaults, model list

Requirements

The LightSaber framework — used for binary streams, app-data paths, file I/O, and logging

Installation

Open LightSaberAI.dpk in Delphi and click Compile.

If you want to know more about how to work with Delphi packages see my book.

Usage

The typical flow looks like this:

var
  Client: TAiClientEx;
  Response: TAIResponse;
  TextPart, SysInstr: TJSONObject;
begin
  Client:= TAiClientEx.Create;  // Auto-loads settings from disk
  try
    Client.UploadFiles(MyInputFiles);

    TextPart:= MakeTextPart('Analyze this image and return findings.');
    SysInstr:= TJSONPair.Create('system_instruction', ...);

    Response:= Client.SendAPICall(TextPart, MyInputFiles, SchemaFilePath, 'debug-name', SysInstr);
    try
      if Response.Valid
      then UseResult(Response.ExtractedJSONObj)
      else ShowError(Response.ErrorMsg);
    finally
      FreeAndNil(Response);
    end;
  finally
    FreeAndNil(TextPart);
    FreeAndNil(SysInstr);
    FreeAndNil(Client);  // Auto-saves settings
  end;
end;