evalytic 0.4.1

Evaluation SDK for AI outputs

Evalytic

Evals for AI outputs. Automated quality evaluation for images, video, text, RAG, and agent runs.

Know if your AI outputs are good before your users tell you they're not. One SDK for visual generation, LLM text, RAG pipelines, and tool-using agents.

pip install evalytic

# Visual: compare image generation models
evaly bench -m flux-schnell -m flux-dev -m flux-pro \
  -p "A product photo on marble countertop" --yes

# RAG: evaluate retrieval-augmented answers
evaly rag eval \
  --query "What is Evalytic?" \
  --response "Evalytic evaluates AI outputs." \
  --context "Evalytic is an evaluation platform." \
  --metrics faithfulness,answer_relevancy

# Text: evaluate LLM outputs
evaly text eval \
  --input "Translate: Hello" --output-text "Merhaba" --expected "Merhaba" \
  --metrics exact_match,semantic_similarity

# Agent: evaluate tool-using agent runs
evaly agent eval \
  --input "Find the score" --final-output "0.9" \
  --tool-call search --expected-tool search

What It Does

Evalytic scores AI outputs across four eval domains with the same consensus-judge architecture:

• Visual (images, video) — 7 semantic dimensions scored by VLM judges + deterministic metrics (sharpness, CLIP, LPIPS, ArcFace, NIMA, TOPIQ)
• RAG — reference-free faithfulness + answer_relevancy; reference-based context_precision + context_recall
• Text — factual_correctness, semantic_similarity, g_eval (custom rubric), BLEU, ROUGE, exact_match, levenshtein, string_presence
• Agent — tool_call_accuracy, goal_accuracy, step_efficiency

VLM / LLM judges (Gemini, GPT, Claude, Ollama) + local metrics work together or independently. Every domain supports consensus mode (2+1 adaptive multi-judge).

Use Cases

• Model Selection — Compare any fal.ai / OpenAI / Anthropic models head-to-head
• RAG Hallucination Detection — Claim-level faithfulness against retrieved context
• Prompt Optimization — Measure output quality across semantic dimensions
• Regression Detection — Catch quality drops when models, prompts, or retrievers update
• CI/CD Quality Gate — Block deploys when any metric falls below threshold (visual OR text OR agent)
• Consensus Judging — Multi-judge scoring with automatic agreement analysis

Quickstart

1. Install

pip install evalytic

2. See Real Examples (no API key needed)

evaly demo              # Opens showcase with 4 real benchmark case studies
evaly demo face         # Face identity preservation comparison
evaly demo flagship     # Flux Schnell vs Dev vs Pro cost/quality

3. Score an Existing Image

# Local metrics only (free, no API key)
evaly eval --image output.png --prompt "A sunset over mountains" --no-judge

# With VLM judge
export GEMINI_API_KEY=your_gemini_key
evaly eval --image output.png --prompt "A sunset over mountains" --yes

4. Benchmark Models

export FAL_KEY=your_fal_key

# Text-to-image
evaly bench -m flux-schnell -m flux-dev -m flux-pro \
  -p "A cat sitting on a windowsill" --yes

# Image-to-image
evaly bench -m flux-kontext -m seedream-edit -m reve-edit \
  --inputs product.jpg -p "Place on a marble countertop" --yes

# Metrics only, no VLM judge
evaly bench -m flux-schnell -m flux-dev -p "A cat" --no-judge

5. Interactive Setup

evaly init   # Guided setup: use case, API keys, config file

CLI Commands

Command	Domain	Description
evaly init	Any	Interactive setup wizard
evaly demo	Visual	Browse real benchmark showcases (no API key needed)
evaly bench	Visual	Generate, score, and report in one command
evaly eval	Visual	Score a single image without generation
evaly rag eval	RAG	Evaluate RAG answers (reference-free + reference-based)
evaly text eval	Text	Evaluate LLM outputs (factual, semantic, BLEU, ROUGE, G-Eval)
evaly agent eval	Agent	Evaluate tool-using agent runs
evaly compare	All	Delta between two report files (same-type only)
evaly gate	All	CI/CD quality gate (--threshold for visual, --metric-threshold for text/RAG/agent)
evaly dataset	All	Manage evaluation datasets (rag, text, agent, visual)

Judges

Any VLM that can analyze images works as a judge:

evaly bench -m flux-schnell -p "A cat" -j gemini-2.5-flash            # Default
evaly bench -m flux-schnell -p "A cat" -j openai/gpt-5.2              # OpenAI
evaly bench -m flux-schnell -p "A cat" -j anthropic/claude-sonnet-4-6 # Anthropic
evaly bench -m flux-schnell -p "A cat" -j fal/gemini-2.5-flash        # Via fal.ai (single key)
evaly bench -m flux-schnell -p "A cat" -j ollama/qwen2.5-vl:7b        # Local

Consensus Mode

Use multiple judges for more reliable scores:

evaly bench -m flux-schnell -p "A cat" \
  --judges "gemini-2.5-flash,openai/gpt-5.2"

Two judges score in parallel. If they disagree, a third breaks the tie.

Optional Extras

pip install "evalytic[metrics]"      # CLIP + LPIPS + ArcFace + NIMA + TOPIQ (~2GB)
pip install "evalytic[ocr]"          # OCR text accuracy (pytesseract)
pip install "evalytic[embeddings]"   # Local sentence-transformers for RAG/text embeddings (~500MB)
pip install "evalytic[all]"          # Everything

For RAG answer_relevancy and text semantic_similarity, either install evalytic[embeddings] or set OPENAI_API_KEY / FAL_KEY (embeddings auto-resolve).

Configuration

Create evalytic.toml in your project root:

[keys]
fal = "your_fal_key"
gemini = "your_gemini_key"

[bench]
judge = "gemini-2.5-flash"
dimensions = ["visual_quality", "prompt_adherence"]
concurrency = 4

[bench.dimension_weights]
input_fidelity = 0.5
visual_quality = 0.1

[rag]
judge = "gemini-2.5-flash"
judges = ["gemini-2.5-flash", "openai/gpt-5.2"]  # Consensus mode

[rag.thresholds]
faithfulness = 0.8
answer_relevancy = 0.7

[text.thresholds]
factual_correctness = 0.8
semantic_similarity = 0.7

[embeddings]
provider = "sentence-transformers"  # or "openai" / "fal"
model = "all-MiniLM-L6-v2"

Documentation

Full docs at docs.evalytic.ai

License

MIT