Aludel gives teams a clean way to evaluate prompt and model behavior without inventing their own tooling first.

• Compare the same prompt across OpenAI, Anthropic, Gemini, Ollama, xAI, Groq, and OpenRouter.
• Inspect output, latency, token usage, and cost side by side.
• Compare prompt versions and see pass-rate, cost, and latency changes over time.
• Run evaluation suites with assertions, document attachments, and CSV or JSON test case imports.
• Execute suites headlessly with machine-readable JSON output for CI workflows.
• Route runs and suites through your app's real LLM workflow with callback execution.
• Reuse single-turn and multi-turn datasets across suites with provenance and metadata filtering.
• Find quality, cost, latency, stability, and regression trade-offs with rolling analytics and Pareto analysis.
• Generate failure-grounded prompt suggestions, then explicitly accept or dismiss them.
• Use it inside an existing Phoenix app or run it standalone.

Why Aludel

Most teams evaluating LLM behavior end up with some combination of scripts, spreadsheets, and ad hoc dashboards. Aludel brings that work into one place with a UI that is practical enough for day-to-day iteration.

• Provider comparison: run the same input across models and vendors in one view.
• Prompt history: keep prompt changes traceable instead of losing them in copy-pasted variants.
• Regression coverage: turn important scenarios into repeatable suites with assertions.
• Embedded app callbacks: evaluate your production-facing workflow without rebuilding it in the dashboard.
• Phoenix-native deployment: mount it in your app or run it as a standalone dashboard.

Feature Catalog

See the complete feature guide for behavior, constraints, and examples.

Structured Output Scoring

Suites support strict string assertions and structured JSON checks.

For structured outputs, use json_deep_compare to score partial matches instead of forcing all-or-nothing pass/fail outcomes.

[
  {
    "type": "json_deep_compare",
    "expected": {
      "status": "ok",
      "customer": {
        "name": "Jane",
        "tier": "gold"
      }
    },
    "threshold": 75.0
  }
]

Aludel stores field-level comparison details, per-test match scores, and suite-run average scores so prompt evolution and exports can track structured output quality over time.

Test Case Imports

Suite pages can import test cases from CSV or JSON. Aludel validates the file and shows a preview with row-level errors before saving any accepted test cases.

• JSON files contain an array of objects with input, expected, and assertion keys.
• CSV files use an input,expected,assertion header row and may include notes.

Quick Start

Embed in an existing Phoenix app

Requirements:

• Elixir and Phoenix
• PostgreSQL 12+

Aludel depends on PostgreSQL-specific features, including JSONB, percentile_disc(), and DATE()-based aggregations. SQLite and MySQL are not supported.

1. Add the dependency

def deps do
  [
    {:aludel, "~> 0.6.1"}
  ]
end

mix deps.get

2. Configure the repo

config :aludel, repo: YourApp.Repo

3. Install and run migrations

mix aludel.install
mix ecto.migrate

4. Mount the dashboard

use YourAppWeb, :router
import Aludel.Web.Router

if Mix.env() == :dev do
  scope "/dev" do
    pipe_through :browser
    aludel_dashboard "/aludel"
  end
end

5. Start using it

Visit your configured path, for example http://localhost:4000/dev/aludel.

Execution modes

Aludel supports two execution modes:

• Native (default): Aludel renders the prompt template and calls the configured provider directly.
• App Callback: your host app executes the real workflow and returns a normalized result back to Aludel.

Use callback mode when your production behavior includes orchestration beyond a single prompt, such as retrieval, tool usage, routing, retries, or post-processing.

Configure it in your embedded app:

config :aludel,
  execution_mode: :callback,
  executor: MyApp.AludelExecutor

Example executor:

defmodule MyApp.AludelExecutor do
  @behaviour Aludel.Executor

  @impl true
  def run(%{
        kind: kind,
        variables: variables,
        documents: documents,
        provider: provider,
        metadata: metadata
      }) do
    case MyApp.AI.reply(%{
           question: variables["question"],
           documents: documents,
           provider: provider && provider.provider,
           model: provider && provider.model,
           context: %{source: :aludel, kind: kind, metadata: metadata}
         }) do
      {:ok, reply} ->
        {:ok,
         %{
           output: reply.text,
           input_tokens: Map.get(reply, :input_tokens),
           output_tokens: Map.get(reply, :output_tokens),
           latency_ms: Map.get(reply, :latency_ms),
           cost_usd: Map.get(reply, :cost_usd),
           metadata: %{trace_id: Map.get(reply, :trace_id)}
         }}

      {:error, reason} ->
        {:error, reason}
    end
  end
end

Success responses only require output. input_tokens, output_tokens, latency_ms, cost_usd, and metadata are optional.

In callback mode, the existing run and suite UI stays the same:

• provider selection still stays available
• the run and suite screens show Execution Mode
• missing token or cost metrics render as N/A
• exports include callback metadata when present
• normalized execution artifacts record the request shape, execution mode, output, metrics, and bounded error details

Native multi-provider runs execute concurrently by default, with a maximum concurrency of three and a 120-second timeout. Hosts can choose sequential execution or tune those limits:

config :aludel,
  run_execution_mode: :concurrent

config :aludel, :llm,
  max_concurrency: 5,
  request_timeout_ms: 120_000

Set run_execution_mode: :sequential when provider calls must not overlap.

Headless suite execution

Run a suite from scripts or CI with stable JSON output:

mix aludel.eval \
  --suite-id SUITE_ID \
  --prompt-version-id PROMPT_VERSION_ID \
  --provider-id PROVIDER_ID

The task exits unsuccessfully when its arguments or targets are invalid, execution cannot complete, the suite has no test cases, or any test case fails.

Successful and failed executions emit a stable aludel_eval JSON envelope with a schema version, suite and provider identifiers, aggregate pass/fail, score, cost and latency data, plus individual assertion results.

Standalone mode

If you want to run Aludel by itself:

git clone https://github.com/ccarvalho-eng/aludel.git
cd aludel/standalone
mix deps.get
mix ecto.create
mix ecto.migrate
mix phx.server

To populate the local database with realistic prompts, providers, datasets, suites, AI-like results, and 60 days of comparison history:

mix aludel.seed

Visit http://localhost:4000.

The standalone release also supports optional HTTP Basic Authentication and read-only access:

export BASIC_AUTH_USER=admin
export BASIC_AUTH_PASS=change-me
export READ_ONLY=true

To smoke-test callback mode in the standalone app, configure a local executor module in standalone/lib/aludel_dash.ex or another module loaded by the standalone app, then add:

config :aludel,
  execution_mode: :callback,
  executor: AludelDash.Executor

After restarting mix phx.server, create a prompt version and provider in the UI, then:

1. Launch a run from /runs/new?version=<prompt_version_id>
2. Run a suite from /suites/<suite_id>
3. Confirm both screens show Execution Mode
4. Confirm the outputs come from your executor and optional metrics render cleanly when omitted

Provider support

Aludel supports OpenAI, Anthropic, Google Gemini, Ollama, xAI, Groq, and OpenRouter.

Provider forms discover active text-generation models from LLMDB, keep deprecated models available when editing existing configurations, and allow custom model IDs. Token costs use built-in per-model rates when available; custom input and output rates can be configured per provider.

For embedded apps, configure provider keys in config/runtime.exs:

# In config/runtime.exs
config :aludel, :llm,
  openai_api_key: System.get_env("OPENAI_API_KEY"),
  anthropic_api_key: System.get_env("ANTHROPIC_API_KEY"),
  google_api_key: System.get_env("GOOGLE_API_KEY"),
  xai_api_key: System.get_env("XAI_API_KEY"),
  groq_api_key: System.get_env("GROQ_API_KEY"),
  openrouter_api_key: System.get_env("OPENROUTER_API_KEY")

Ollama runs locally and does not require an API key.

Callback mode does not require Aludel to use those API keys directly, but provider selection still remains part of the current run and suite flows and is passed into the executor for host-app routing when needed.

Document Storage

Uploaded test case documents go through Aludel.Storage. Documents can be attached while creating new suite test cases or while editing existing test cases.

Supported uploads are PDF, PNG, JPEG, JSON, CSV, and plain text. Anthropic accepts PDFs natively; adapters that require images can use the configurable ImageMagick converter.

• Development uses the local filesystem adapter from config/dev.exs.
• Production uses config/runtime.exs and requires ALUDEL_STORAGE_BACKEND.

Development storage

Development stores uploaded documents on the local filesystem.

Production storage

Set ALUDEL_STORAGE_BACKEND to aws or gcs.

For AWS S3:

export ALUDEL_STORAGE_BACKEND=aws
export AWS_S3_BUCKET=aludel-uploads
export AWS_REGION=us-east-1
export AWS_ACCESS_KEY_ID=...
export AWS_SECRET_ACCESS_KEY=...

For Google Cloud Storage:

export ALUDEL_STORAGE_BACKEND=gcs
export GCS_BUCKET=aludel-uploads
export GOOGLE_APPLICATION_CREDENTIALS=/absolute/path/to/service-account.json

If your GCS bucket requires requester-pays access, also set:

export GCS_USER_PROJECT=your-billing-project-id

The GCS adapter uses Goth with standard Google application credentials. GOOGLE_APPLICATION_CREDENTIALS_JSON also works if you prefer inline JSON.

Evaluation Workflows

Test cases can be authored in the visual assertion editor or as JSON. Suites support inline editing, document management, execution history, detailed assertion results, and retrying an individual failed result without rerunning the entire suite.

Reusable datasets hold ordered single-turn or multi-turn entries. Each entry can include template variables, conversation messages, assertions, and arbitrary JSON metadata. Dataset pages support metadata filtering, and suite population preserves source provenance while skipping entries that were already imported into that suite.

Prompt evolution combines suite history across versions and providers. The UI shows pass rate, structured-output score, cost, latency, version-over-version deltas, regression and stability signals, and a suite-scoped Pareto frontier. A failure reflection workflow can ask the selected provider for a variable-preserving prompt suggestion; accepting it creates a new immutable prompt version, while dismissal keeps the decision in history.

Documentation

The README is intentionally optimized for first contact. For deeper setup, usage, and contribution details:

• Feature Guide
• Evaluation Guide
• Embedding Guide
• Wiki
• HexDocs
• Contributing Guide
• Issue Tracker
• Discussions

Development

For local development:

mix deps.get
mix compile
mix test
mix precommit

If you are changing frontend assets:

mix assets.build
mix compile --force

For standalone development, run the app from the standalone directory:

cd standalone
mix phx.server

If you change frontend assets, rebuild them from the repo root and restart the standalone server:

mix assets.build
mix compile --force

License

Apache License 2.0