substrai-evalforge 0.4.0

Automated LLM evaluation pipeline generator

EvalForge

Automated LLM evaluation pipeline generator.

Built by SubstrAI — Open-source GenAI frameworks for serverless infrastructure.

The Problem

Every team deploying LLMs builds evaluation pipelines from scratch. RAGAS and DeepEval are libraries — they don't generate infrastructure, schedule runs, detect drift, or route to human reviewers.

The Solution

Describe your use case → EvalForge generates the complete evaluation pipeline:

# evalforge.yaml
use_case:
  type: rag
  description: "Customer support chatbot"
  model:
    provider: bedrock
    model_id: anthropic.claude-3-haiku-20240307-v1:0

evaluation:
  metrics: auto  # auto-selects: faithfulness, relevancy, precision, recall, toxicity

evalforge run
# Faithfulness:      0.91 ✓ (threshold: 0.85)
# Answer Relevancy:  0.87 ✓ (threshold: 0.80)
# Context Precision: 0.78 ✓ (threshold: 0.75)
# Toxicity:          0.02 ✓ (threshold: 0.05)
# Overall: PASS (4/4 metrics passing)

Features

• Use-case-driven metric selection — describe your app, get optimal metrics
• 6 use case types — RAG, summarization, classification, generation, chat, code
• 16+ built-in metrics — faithfulness, ROUGE, BLEU, toxicity, injection resistance, F1
• Synthetic test data generation — adversarial, edge cases, domain-specific
• Drift detection — alerts when quality degrades over time
• Human-in-the-loop — route uncertain evaluations to reviewers
• Scheduled pipelines — daily/weekly automated evaluation runs
• Benchmark registry — compare against published benchmarks
• One-command deploy — Step Functions + Lambda infrastructure

Installation

pip install substrai-evalforge

Quick Start

# Scaffold project
evalforge init my-eval --use-case rag

# Run evaluation
cd my-eval
evalforge run

# List available metrics
evalforge metrics --use-case rag

Python SDK

from evalforge import EvalPipeline

# Quick start for any use case
pipeline = EvalPipeline.for_use_case("rag")
results = pipeline.run()
print(results.summary())
print(f"All passing: {results.all_passing}")

Supported Use Cases & Auto-Selected Metrics

Use Case	Auto-Selected Metrics
rag	faithfulness, answer_relevancy, context_precision, context_recall, toxicity
summarization	rouge_l, bleu, coherence, conciseness, fluency
classification	accuracy, precision, recall, f1_score
generation	fluency, coherence, toxicity, bias_detection
chat	coherence, toxicity, injection_resistance, fluency
code	accuracy, coherence

License

MIT — see LICENSE

Author

Gaurav Kumar Sinha — Founder, SubstrAI

• Email: [e-mail]
• GitHub: @substrai