LLM Eval Workbench

From medieval Latin, derived from Arabic al-uthāl — a vessel for sublimation, where matter is refined through stages.
Like jinn answering invocation, LLMs respond to prompts; here, their nature is revealed, tested, and distilled.

Run prompts across OpenAI, Anthropic, and Ollama simultaneously. Compare output quality, latency, token usage, and cost in real-time.

Features

Multi-provider comparison — Run the same prompt across providers side-by-side. Track latency, token usage, and cost per run.
Prompt management — Version-controlled templates with {{variable}} interpolation. Every edit creates an immutable new version. Supports tags and descriptions.
Evolution tracking — Visualize prompt version performance over time. Track pass rates, cost, and latency trends across versions and providers.
Evaluation suites — Visual test case editor with document attachments (PDF, images, CSV, JSON, TXT). Automated assertions including contains, regex, exact_match, and json_field. Track pass rates and catch regressions over time.
Dashboard — Live metrics as runs execute: cost trends, latency, and per-provider performance.

Installation

Aludel can be embedded into any Phoenix LiveView application as a self-contained dashboard.

1. Add dependency

Add Aludel to your mix.exs:

def deps do
  [
    {:aludel, "~> 0.1"}
  ]
end

Run mix deps.get

2. Configure the repo

Add to config/config.exs:

config :aludel, repo: YourApp.Repo

3. Install migrations

mix aludel.install

This copies Aludel’s migrations to your priv/repo/migrations/ directory.

4. Run migrations

mix ecto.migrate

5. Add router macro

In your lib/your_app_web/router.ex:

use YourAppWeb, :router
import Aludel.Web.Router  # Add this line

# In development
if Mix.env() == :dev do
  scope "/dev" do
    pipe_through :browser
    aludel_dashboard "/aludel"  # Dashboard will be at /dev/aludel
  end
end

# Or in production (with authentication)
# scope "/admin" do
#   pipe_through [:browser, :require_admin]
#   aludel_dashboard "/aludel"
# end

The dashboard can be mounted at any path you choose. It’s common to mount it under /dev in development or /admin in production (with proper authentication).

6. Configure API keys (optional)

Aludel reads provider API keys from application config. Add to your host app’s config:

# config/dev.exs (or config/runtime.exs for production)
config :aludel, :llm,
  openai_api_key: System.get_env("OPENAI_API_KEY"),
  anthropic_api_key: System.get_env("ANTHROPIC_API_KEY")

Then set environment variables before starting the server:

export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...
mix phx.server

Ollama runs locally and requires no API keys.

7. Install ImageMagick (for Ollama PDF support)

If you want to test evaluation suites with PDF documents using Ollama vision models, install ImageMagick v7+:

# macOS
brew install imagemagick

# Ubuntu/Debian
sudo apt-get install imagemagick

# Check installation
magick -version

Note: PDF-to-image conversion is only required for Ollama vision models. OpenAI and Anthropic Claude 4.5+ accept PDFs directly in their APIs without conversion. For Ollama, PDFs are converted to PNG (first page only, 150 DPI) before being sent to the model.

8. Seed demo data (optional)

mix aludel.seed

This populates the database with sample providers, prompts, and evaluation suites.

Visit the dashboard at your configured path (e.g., http://localhost:4000/dev/aludel).

Standalone Mode

Aludel includes a standalone application in the standalone/ directory for running the dashboard without embedding it in a Phoenix app.

Setup

cd standalone
mix deps.get
mix ecto.create
mix ecto.migrate
mix aludel.seed  # Optional: add demo data
mix phx.server

Visit http://localhost:4000

Configuration

Edit standalone/config/dev.exs to configure:

Database — Default: postgres://postgres:postgres@localhost/aludel_dash_dev
Port — Default: 4000
API Keys — Set OPENAI_API_KEY and ANTHROPIC_API_KEY environment variables

Production Deployment

# Set required environment variables
export DATABASE_URL=postgres://...
export SECRET_KEY_BASE=$(mix phx.gen.secret)
export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...

# Optional: Enable basic auth
export BASIC_AUTH_USER=admin
export BASIC_AUTH_PASS=secret

# Optional: Set read-only mode
export READ_ONLY=true

# Run the app
MIX_ENV=prod mix release
_build/prod/rel/aludel_dash/bin/aludel_dash start

Providers

Provider	API Key Required	Configuration
Ollama ⭐	No	Local models - works out of the box
OpenAI	Yes	Add `OPENAI_API_KEY` to `.env` (Get key)
Anthropic	Yes	Add `ANTHROPIC_API_KEY` to `.env` (Get key)

Ollama quickstart

# Install from https://ollama.com, then:
ollama serve
ollama pull llama3  # or: mistral, codellama
mix run priv/repo/seeds.exs

Seeds create providers for Ollama, OpenAI, and Anthropic, along with 3 sample prompts and 3 evaluation suites with document attachments.

Usage

1. Create a prompt — Go to Prompts → New Prompt and use {{variable}} syntax:

Explain {{topic}} in exactly 3 sentences.

2. Run across providers — Click New Run, fill in variables, select providers, and watch results stream in.

3. Build evaluation suites — Go to Suites → New Suite, add test cases with assertions, and run regression tests.

4. Track evolution — View the Evolution tab on any prompt to see how versions improve over time. Metrics show pass rates, cost, and latency per version and provider.

5. Add a provider — Go to Providers → New Provider, select the type, choose a model, and configure parameters (temperature, max_tokens, etc.). API keys are set via environment variables in .env.

Development

Working with assets (CSS/JS)

Aludel uses Tailwind CSS and esbuild for styling and JavaScript bundling.

After making changes to CSS or JS files:

# 1. Rebuild assets (from aludel directory)
mix assets.build

# 2. Force recompile to pick up new asset hashes
mix compile --force

# 3. If working on an embedded installation, recompile the dependency
cd ../your_host_app
mix deps.compile aludel --force

# 4. Restart the Phoenix server to pick up changes

Asset files:

CSS: assets/css/app.css
JavaScript: assets/js/app.js and assets/js/hooks/
Built assets: priv/static/app.css and priv/static/app.js (committed to git)

Live development workflow:

For faster iteration during development, you can use Mix tasks with watchers:

# Watch and rebuild CSS on changes
mix tailwind aludel --watch

# Watch and rebuild JS on changes (in another terminal)
mix esbuild aludel --watch

Alternatively, run the standalone app for a full development server:

cd standalone
mix phx.server  # Starts asset watchers automatically

Community

💬 Discussions — Ask questions, share ideas, or discuss use cases
🐛 Issues — Report bugs or request features

Contributing

See CONTRIBUTING.md for detailed guidelines.

Quick start:

Fork the repository
Create a feature branch (git checkout -b feature/your-feature)
Commit using conventional commits
Run mix precommit before submitting
Open a Pull Request

License

Apache License 2.0