Aludel - LLM Evaluation and Prompt Testing for Elixir

Aludel is a Phoenix-native LLM evaluation workbench for Elixir applications. It combines prompt testing, model comparison, regression suites, red-team datasets, quality gates, and LLM observability in an embeddable dashboard and standalone release.

• 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 deterministic assertions, model-based judges, repeated sampling, document attachments, and CSV or JSON test case imports.
• Assert generated output directly in ExUnit, or execute and gate persisted suites from application tests and CI.
• Execute suites headlessly from versioned JSON or YAML manifests with console, JSON, JUnit XML, or GitHub annotation reports.
• 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.
• Materialize versioned prompt-injection, data-disclosure, unsafe-action, and misinformation checks from a curated red-team catalog, or generate bounded candidates for explicit review.
• 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.

UI, CLI, and Library APIs

The dashboard and automation interfaces use the same persisted prompts, providers, datasets, suites, runs, quality policies, and evaluation evidence. A workflow can be authored in the UI, committed as a suite manifest, gated from the CLI, and inspected again in the dashboard without maintaining a second test-case format.

For direct integration, start with the Aludel.Datasets, Aludel.Evals, Aludel.Prompts, Aludel.Prompts.Optimization, Aludel.Stats.Overview, Aludel.Storage, Aludel.Execution, and Aludel.Executor API references.

Feature Catalog

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

Model-Based Evaluation

Use a rubric judge when correctness depends on meaning instead of an exact string or JSON shape. Choose a separate configured provider for the judge and use either a custom rubric or a versioned built-in template:

[
  {
    "type": "rubric_judge",
    "template": "faithfulness",
    "provider_id": "00000000-0000-0000-0000-000000000000",
    "threshold": 85,
    "context": "Claims must be supported by this grounding evidence."
  }
]

Replace the placeholder with the ID of a configured provider.

In the dashboard, add or edit a suite test case, choose rubric judge in the visual assertion editor, then select a built-in judge or enter a custom rubric. The same controls set the judge provider, pass threshold, optional reference answer, and optional grounding context. JSON assertions remain available for direct editing, and persisted judge assertions run unchanged from the dashboard, mix aludel.eval, ExUnit, and the Elixir API.

The catalog includes correctness, relevance, faithfulness, safety, refusal quality, PII protection, and hallucination checks. Aludel records the resolved rubric and template version alongside score, reasoning, duration, provider, model, token usage, cost, and structured evaluator status.

See the evaluation guide, rubric judge guide, and judge catalog.

Suite execution supplies metrics with normalized output, rendered input, prompt template, variables, messages, documents, metadata, provider, prompt version, and execution evidence. Expected references remain available in assertion configuration, and direct callers can also set expected on Aludel.Evals.Metric.Context. See metric context and evaluator execution details.

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.

Repeated Evaluation Sampling

Nondeterministic model output can make a single response misleading. Run each test case up to 20 times and reduce the ordered attempts into one result:

In the dashboard, open a suite and set Attempts per test case plus the Pass rule before running it. Choose all attempts, any attempt, a strict majority, or a minimum pass-rate percentage. Each sampled result expands to show the aggregate decision and every ordered attempt.

{:ok, suite_run} =
  Aludel.Evals.execute_suite(suite, prompt_version, provider,
    samples: 5,
    reducer: {:minimum_pass_rate, 0.8}
  )

Reducers support :all, :any, strict :majority, and a minimum pass rate. Aludel retains every attempt, sums token usage, cost, and latency, and reruns the complete sampling configuration when a result is retried.

Dashboard percentages map to the API's 0.0–1.0 minimum pass rate. File-based suites expose the same configuration to mix aludel.eval, while the Elixir API and ExUnit accept the sampling options directly.

See the evaluation guide and repeated sampling wiki guide for the full result shape and reducer examples.

Curated Red-Team Datasets

Materialize a versioned set of adversarial cases into any reusable dataset:

In the dashboard, open a dataset and choose Add red-team cases. Browse every catalog prompt, risk reference, severity, technique, deterministic canary, and recommended judge before selecting exact cases. Choose the target prompt variable and optionally a configured judge provider and threshold; the result reports newly created and already-present entries.

Choose Generate cases for product-specific candidates. Configure explicit request, token, cost, and timeout bounds; inspect the receipt, safe failures, usage, checksums, prompt, rationale, classification, and recommended judge; then check the exact candidates to import. Nothing is pre-approved, and unimported candidates are discarded when the connected page session ends.

{:ok, dataset} = Aludel.Datasets.create_dataset(%{name: "Security regressions"})

{:ok, %{created: created, skipped: skipped}} =
  Aludel.RedTeam.materialize(dataset,
    categories: [:prompt_injection, :system_prompt_leakage],
    judge_provider_id: judge_provider.id,
    judge_threshold: 90
  )

Each curated entry includes a deterministic canary assertion plus optional model-based judging. Metadata records its stable case ID, catalog and case versions, risk category, severity, technique, source, checksum, and deduplication key. Repeating the same materialization skips matching entries; conflicting content or judge configuration returns an error.

Generate cases for a product-specific context without writing to the database, then inspect the returned candidates and failures before deciding which cases belong in an evaluation dataset:

{:ok, generation} =
  Aludel.RedTeam.generate(generator_provider.id,
    categories: [:prompt_injection, :sensitive_information_disclosure],
    target_context: "A support assistant may read account history but must not reveal credentials",
    cases_per_category: 2,
    max_requests: 2,
    max_total_tokens: 8_000,
    max_cost_usd: 1.00
  )

Generation is bounded by category count, cases per category, context size, requests, output tokens, observed total tokens, observed cost, response size, and per-request timeout. Valid categories remain reviewable when another category fails. Each candidate and the complete generation carry checksums for stable review. Usage is observed from successful provider responses; failed or timed-out external requests may still incur provider-side usage that Aludel cannot report.

After reviewing the candidates, explicitly approve their stable IDs and import them atomically:

approved_case_ids =
  generation.cases
  |> Enum.filter(&approved_by_reviewer?/1)
  |> Enum.map(& &1.id)

{:ok, %{created: created, skipped: skipped}} =
  Aludel.RedTeam.import_generated(dataset, generation,
    approved_case_ids: approved_case_ids
  )

Import revalidates the generation and candidate checksums, requires at least one unique approved ID, attaches each candidate's recommended rubric judge, records generation and review provenance, and skips only an exact prior import. Any conflict rolls back the complete approved selection.

Curated catalog materialization and generated-case generation, review, and approval/import are available in the dashboard and Elixir API. Persisted cases use the normal dataset and suite workflows in the dashboard, mix aludel.eval, ExUnit, and the library API. There is no separate red-team CLI command.

See the red-team guide, curated datasets wiki guide, and generated cases wiki guide for category filters, review, budgets, provenance, and rerun behavior.

Versioned Quality Policies

Define what a suite must satisfy instead of treating every run as an all-or-nothing test count:

In the dashboard, open a suite and choose Manage policy. The editor validates the JSON definition before saving, shows every immutable version, and uses the newest version for future runs. Suite results identify the snapshotted policy version and show the measured status, actual value, and requirement for every rule.

{:ok, policy} =
  Aludel.Evals.create_suite_policy(suite, %{
    "schema_version" => 1,
    "rules" => [
      %{"id" => "overall", "type" => "overall_pass_rate", "minimum" => 0.95},
      %{
        "id" => "priority",
        "type" => "metadata_pass_rate",
        "metadata" => %{"priority" => "high"},
        "minimum" => 1.0
      },
      %{"id" => "budget", "type" => "total_cost_usd", "maximum" => 0.50}
    ]
  })

Policies can gate overall pass rate, metadata groups, evaluator scores, total cost, and average latency. Each update creates an immutable suite-local version. A run snapshots the latest version before execution, and retries continue to use that same version.

Policy authoring and result inspection are available in the dashboard and through the Elixir API. mix aludel.eval, ExUnit gates, reporters, and exports consume the same stored policy outcome. Results report passed, failed, invalid, or unavailable; missing evidence never silently passes. Suites without a policy preserve the legacy all-cases-must-pass behavior.

See the evaluation guide and quality policies wiki guide for every rule and result example.

Evaluation Reporters

Render the same persisted suite result for local review, API consumers, and CI systems without coupling those formats to persistence:

In the dashboard, open a suite result and choose Reports to preview or download console text, schema-version-2 JSON, JUnit XML, or GitHub Actions annotations. Browser previews are bounded, downloads are not cached, and JUnit excludes generated responses unless Include generated output is explicitly enabled. The existing full run export remains available for complete persisted evidence.

alias Aludel.Evals.Reporter

console = Reporter.render!(suite_run, :console)
json = Reporter.render!(suite_run, :json, pretty: true)
junit_xml = Reporter.render!(suite_run, :junit)
github_annotations = Reporter.render!(suite_run, :github)

Aludel.Evals.Report defines the stable schema-version-2 model. Built-in reporters cover concise console text, JSON interchange, JUnit test reports, and GitHub Actions annotations. Custom modules can implement the Aludel.Evals.Reporter behavior.

The dashboard, Mix CLI, and Elixir API expose the same four built-in formats. The GitHub reporter escapes and bounds untrusted evaluator text before emitting workflow commands. JUnit output escapes identifiers and reasons, omits model output by default, and represents a policy-only rejection as a failed test case. Pass include_output: true, enable it in the dashboard, or use --include-output only when the target artifact is appropriate for generated responses.

See the reporter guide and evaluation reporters wiki guide for CLI, library, CI, and custom reporter examples.

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, "~> 0.7.0"}
  ]
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

Keep the execution target and sampling policy in a versioned manifest while the suite's test cases and dataset provenance remain in Aludel:

schema_version: 1
suite_id: 9a756a58-eaec-43ca-99e6-f5c016d85d0c
prompt_version_id: e74cf2e1-94b6-4bcb-9ed9-b259661be906
provider_id: e1c60ec0-6d55-419b-b958-7d088055254f
sampling:
  samples: 5
  reducer: majority

Run it from scripts or CI:

mix aludel.eval --file evals/support-answer.yaml

JSON manifests use the same schema. You can also supply the three identifiers directly:

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

Select another reporter and optionally write it directly to a file:

mix aludel.eval \
  --suite-id SUITE_ID \
  --prompt-version-id PROMPT_VERSION_ID \
  --provider-id PROVIDER_ID \
  --format junit \
  --output aludel-junit.xml

Supported formats are console, json, junit, and github; --pretty formats JSON for human review, while --include-output opts JUnit into generated responses. The task exits unsuccessfully when its arguments or targets are invalid, execution cannot complete, the suite is empty, or the active quality gate does not pass.

The versioned aludel_eval JSON envelope includes suite and provider identifiers, aggregate and total score/cost/latency data, quality-policy evidence, and individual assertion results.

See the file-based suite guide and wiki examples for both formats, every sampling reducer, library execution, validation behavior, and CI usage.

ExUnit evaluation assertions

Use Aludel.ExUnit to keep focused model checks beside application tests:

defmodule MyApp.AnswerTest do
  use ExUnit.Case
  use Aludel.ExUnit

  test "answers with the expected city" do
    output = MyApp.answer("What is the capital of France?")

    assert_evaluation(output, %{
      "type" => "exact_match",
      "value" => "Paris"
    })
  end
end

Use assert_evaluations/2 for several metrics, assert_suite_run/1 for an existing persisted result, or assert_suite/3 and assert_suite/4 to execute, persist, and gate a suite. Stored quality policies determine the effective suite status when present; otherwise every case must pass and empty runs fail. Failure messages include bounded metric, case, and policy evidence without copying generated output into test logs.

See the ExUnit evaluation guide and ExUnit wiki examples.

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 requires HTTP Basic Authentication in production and also supports read-only access:

export BASIC_AUTH_USER=admin
export BASIC_AUTH_PASS="$(openssl rand -base64 32)"
export READ_ONLY=true

Production startup rejects missing, partial, or blank credentials. Local development remains unauthenticated and binds only to loopback. Serve production Basic Authentication over TLS; if a reverse proxy terminates TLS, preserve the Authorization header and do not expose the backend port directly.

Docker Compose deployments also require a generated POSTGRES_PASSWORD. The database stays private to the Compose network and the web release receives separate connection fields, so passwords containing URI-reserved characters work without encoding. Existing Compose volumes need a one-time password rotation before upgrade; see the embedding and deployment guide.

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")

Provider credentials are runtime-only. Provider configuration accepts generation settings such as temperature and max_tokens, but rejects credential keys at every nesting level. The database constraint enforces the same rule for direct writes, and the migration removes credential keys from legacy provider configuration.

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.
• Standalone production validates ALUDEL_STORAGE_BACKEND and its backend-specific settings at startup.

Development storage

Development stores uploaded documents on the local filesystem.

Production storage

Set ALUDEL_STORAGE_BACKEND to local, aws, or gcs.

For a persistent local volume:

export ALUDEL_STORAGE_BACKEND=local
export ALUDEL_STORAGE_PATH=/data/aludel_uploads

For AWS S3:

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

The standalone release uses the AWS runtime identity provider when credentials are not explicitly set. When static or temporary credentials are needed, set the access-key pair together; the session token is optional:

export AWS_ACCESS_KEY_ID=...
export AWS_SECRET_ACCESS_KEY=...
export AWS_SESSION_TOKEN=...

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.

Startup rejects missing backend settings, unsupported backends, and partial or blank explicit AWS credentials. Embedded applications can configure Aludel.Storage directly for these adapters or a custom backend.

Changing the active backend does not move existing objects. Keep the former backend variables configured until its documents have been removed or migrated; the standalone release retains every completely configured backend for reads and cleanup of existing document rows.

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
• Reporter 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