Table of Contents

Samples

The repository ships with runnable sample applications that exercise the main integration paths.

Console sample

Project: samples/Hearth.Samples.Console

dotnet run --project samples/Hearth.Samples.Console -- ./models/qwen2.5-7b-q4_k_m.gguf

You can also provide the model path through HEARTH_MODEL.

The console sample is useful for:

  • quick sanity checks against a local model
  • streaming output experiments
  • prompt iteration without building a full app

Blazor sample

Project: samples/Hearth.Samples.Blazor

dotnet run --project samples/Hearth.Samples.Blazor -- ./models/qwen2.5-7b-q4_k_m.gguf

Open http://localhost:5000 after startup. The sample streams tokens into the UI as they are generated using the Hearth.AI.Blazor component library.

Inference server

Project: samples/Hearth.Samples.Server

A minimal ASP.NET Core app that exposes OpenAI-compatible /v1 endpoints. This is the image that Aspire pulls when you use Hearth.AI.Aspire.Hosting. It can also be run standalone:

HEARTH__MODEL=Qwen/Qwen2.5-7B-Instruct-GGUF dotnet run --project samples/Hearth.Samples.Server

The server is also published as a container image:

docker run -p 5000:5000 \
  -e HEARTH__MODEL=Qwen/Qwen2.5-7B-Instruct-GGUF \
  ghcr.io/muneeb-devp/hearth-server:latest

All configuration is via environment variables:

Variable Default Description
HEARTH__MODEL (required) HuggingFace repo ID or local path
HEARTH__GPULAYERS 0 GPU layers to offload
HEARTH__CONTEXTSIZE 4096 Context window size
HEARTH__BATCHSIZE 512 Prompt-processing batch size
HEARTH__HUGGINGFACETOKEN (optional) Token for gated model repos

What to study in the samples

  • how AddHearth(...) is registered
  • how IChatClient is injected
  • how streaming responses are consumed
  • how the app stays independent from the underlying inference host
Edit this page