GenAI_HostedInternal — Local LLM Gateway Evaluation (Delphi)

A small Delphi console project to test connectivity, authentication, and basic requests against a locally hosted / internal LLM gateway that exposes OpenAI-compatible endpoints (e.g., /chat/completions, /models, /embeddings) [2].

---

Purpose

• Evaluate access to locally hosted AI models via a gateway (“LLM gateway console test”) [1]
• Validate:
  • Base URL + endpoint paths (chat, models, embeddings) [2]
  • Authentication/token retrieval (/auth/token) [1] [2]
  • Client configuration: model name, max tokens, timeouts, temperature [1] [2]

---

Project structure

Entry point (console app)

• GenAI_HostedInternal.dpr
  • Prints key runtime settings:
    • Parameter file name
    • “OpenAI base” URL composed from HostURL + CHAT_COMPLETIONS
    • Token endpoint URL HostURL + AUTH_TOKEN_ENDPOINT
    • Selected model and max tokens [1]
  • Includes the key units for communication, token retrieval, and helpers [1]
  • Supports retrieving a token via:
    • Basic auth (username/password) using GetPermanentTokenViaBasic(...)
    • Token endpoint using GetCompanyInternalToken [1]

Configuration

• Config_AIModels.ini
  • Defines the gateway and endpoint paths [2]:
    • BaseURL
    • ChatCompletions, Completions, Embeddings, ImagesGenerations, Models
    • AuthTokenEndpoint
  • Auth section supports Method=Basic or Method=Bearer, plus optional Token [2]
  • TLS section: CABundle, VerifyPeer [2]
  • Client defaults: Model, TimeoutMS, Temperature, MaxTokens, Stream [2]

Core units

• Unit_AIModel.TAIModelConfig.pas

  • Reads config values like ModelsPath and AuthTokenEndpoint from INI [3]
  • Creates HTTP client (TIdHTTP) with JSON headers/timeouts [3]
  • Notes TLS verification is “basic” and suggests production improvements (peer verification / CA handling) [3]

• Unit_AIModel.GetInternalToken.pas

  • Implements GetCompanyInternalToken(...)
  • Default overload calls the parameterized version with timeout and '/auth/token' [7]

• Unit_AIModel.Probing.pas

  • Builds an Authorization header; can auto-fetch token via GetCompanyInternalToken [6]
  • Probes /models via GET BaseURL + ModelsPath and expects HTTP 200 [6]

• Unit_AIModel.Embeddings.pas

  • Builds an embeddings request using Config.BaseURL + Config.EmbeddingsPath [5]
  • Sends OpenAI-style JSON with model, input, and encoding_format=float [5]
  • Contains a simple cache with HIT/MISS messages when debug is enabled [5]

• Unit_AIModel.Communication.pas

  • Provides Basic auth header encoding and request debugging helpers [9]
  • Defines AUTH_TOKEN_ENDPOINT = '/auth/token' [9]

• Unit_AIModel.Helper.pas

  • Contains utilities such as JoinUrl(...) and model-family helpers [8]

---

How to use

1) Configure your gateway URL and endpoints

Edit Config_AIModels.ini:

[server]
BaseURL=https://your-gateway.example
ChatCompletions=/chat/completions
Models=/models
Embeddings=/embeddings
AuthTokenEndpoint=/auth/token