🤖 GenAI Output Validation

💡 Problem

You've built/evaluating your own custom LLM (due to data privacy or confidentiality reasons), or you want to evaluate the performance of a new LLM of your interest against a mature LLM such as ChatGPT, Gemini, LLaMA, etc.

Validating AI output means measuring how accurate, relevant, or reliable an AI-generated response is for a given input prompt (from your LLM of interest), compared to a reference or ideal answer.

Using a JudgeLLM (a mature LLM that is widely accepted, such as ChatGPT, Gemini, LLaMA, etc.) allows you to automate this evaluation.

📊 About Quantitative Validation

Quantitative validation includes:

• 🔍 Comparison against a reference or ground-truth answer.
• 🧮 Scoring AI-generated outputs based on measurable quality metrics.
• 📈 Aggregation of scores to produce an overall performance rating.
• 🧠 Reasoned Justification explaining why a particular score or judgment was given.

---

⚙️🛠️ Solution

flowchart TD

    %% -----------------------------
    %% Core Test Case Structure
    %% -----------------------------
    subgraph TC["LLM Test Case"]
        P["🧩 input_prompt"]
        EA["📘 expected_answer"]
        AGA["🤖 ai_generated_answer"]
        SoT["📄 source_of_truth"]
        AiT["🧠 ai_inferred_truth"]
    end

    %% -----------------------------
    %% Metrics Section
    %% -----------------------------
    subgraph METRICS["GenAI QA Framework"]
        AR["Answer Relevancy"]
        FA["Faithfulness"]
        CP["Contextual Precision"]
        CR["Contextual Recall"]
        CRe["Contextual Relevancy"]
        HA["Hallucination"]
        SU["Summarization"]
    end

    %% -----------------------------
    %% Judge LLM
    %% -----------------------------
    subgraph LLM["Judge LLM"]
        J["AmazonBedrockModel (amazon.nova-pro-v1:0)"]
    end

    %% -----------------------------
    %% Main Flow
    %% -----------------------------
    TC -->|"Passed to"| METRICS
    METRICS -->|"Each metric uses"| LLM
    LLM -->|"Evaluates each test metric for a given LLM Test Case"| METRICS

    %% -----------------------------
    %% Relationships and Outputs
    %% -----------------------------
    METRICS -->|"Generate test scores + reasons"| RES["📊 Metric Results per Test Case"]
    RES -->|"Aggregated across test cases"| AVG["📈 Aggregated Test Metrics"]

    %% -----------------------------
    %% Visual Group Labels
    %% -----------------------------
    classDef main fill:#f9f9f9,stroke:#333,stroke-width:1px;
    classDef block fill:#eef6ff,stroke:#4a90e2,stroke-width:1px;
    classDef judge fill:#fff2cc,stroke:#e6b800,stroke-width:1px;
    classDef result fill:#e6ffed,stroke:#27ae60,stroke-width:1px;

    class TC,SoT,AiT,P,EA,AGA block
    class METRICS,AR,FA,CP,CR,CRe,HA,SU main
    class LLM,J judge
    class RES,AVG result

Loading

---

📸 Demo

---

⚙️ Tech Stack

• Python3 – Core implementation language
• DeepEval – Framework for extensible GenAI test metrics and customizable Judge LLM integration
• AWS Bedrock (Optional) – Used with Amazon Nova Pro as the Judge LLM (for this demo)
• Vega (with altair) – Visualization library for analyzing and illustrating GenAI testing metrics

---

📦 Installation Guide

🖥️ Test Environment

This program was tested using Python 3.10.16 on an Apple M1 Mac running macOS Ventura 13.7.1.

python3 --version
# Python 3.10.16

pip3 --version
# pip 23.0.1 from /Users/{your_username}/.pyenv/versions/3.10.16/lib/python3.10/site-packages/pip (python 3.10)

uname -a
# Darwin {your_machine_name} 22.6.0 Darwin Kernel Version 22.6.0: Thu Sep  5 20:47:01 PDT 2024; root:xnu-8796.141.3.708.1~1/RELEASE_ARM64_T6000 arm64

sw_vers
# ProductName:            macOS
# ProductVersion:         13.7.1
# BuildVersion:           22H221

uname -m
# arm64

sysctl -n machdep.cpu.brand_string
# Apple M1 Max

🧩 Pre-requisites

• 🐍 Install Python3
• ☁️ Ensure you have an AWS account and valid AWS credentials 🔑
• ⚙️ Configure your AWS credentials to your default AWS profile using aws configure command inside the $HOME/.aws folder 📂.

🚀 How to Run

# Clone this repository
git clone https://github.com/tech-magic/gen-ai-validator.git
cd gen-ai-validator

# Create your own python virtual environment
python3 -m venv llm-testing-venv
source llm-testing-venv/bin/activate

# Install all requirements into the python virtual environment
pip install -r requirements.txt

# Run the app from the python virtual environment
python main.py

📝 Notes

🧠 AWS is not a mandatory requirement to run deepeval test cases. It is ONLY listed as a pre-requisite here since Amazon-Nova-Pro is being used as the JudgeLLM for this demo (see modules/judge_llm.py)

You can choose any trusted LLM 🤖 as your JudgeLLM from the following list in the deepeval library:

https://github.com/confident-ai/deepeval/tree/main/deepeval/models/llms

🧠 The available options for choosing a JudgeLLM (as per the above link) includes:

• 💬 OpenAI
• 🔍 Deepseek
• 🌐 Gemini
• ☁️ AWS
• 🔷 Azure
• 🧠 Ollama
• ➕ and more...

💡 You can even create your own custom JudgeLLM 🧩 by extending the class: deepeval.models.DeepEvalBaseModel 🧱

---

🧩 LLM Test Case Structure

Each test case contains the following fields:

Key	Type	Required	Description
input_prompt	str	✅	The question or instruction given to the AI model.
expected_answer	str	✅	The correct or reference answer (ground truth).
ai_generated_answer	str	✅	The actual response produced by the AI model.
test_case_group	str	❌	Optional logical grouping (e.g., “General Knowledge”, “Math”, “Geography”). Defaults to "default".
test_case_name	str	❌	Optional unique name for the test case. Defaults to the input_prompt value.
background_context	str	❌	Optional textual context or reference document supporting the expected answer. If not provided, it is auto-generated as a combination of input_prompt and expected_answer.
ai_inferred_context	str	❌	Optional textual context inferred from the AI’s reasoning. If not provided, it is auto-generated as a combination of input_prompt and ai_generated_answer.

---

🧮 GenAI Test Metrics

All metrics produce a score between 0.0 and 1.0, where 1.0 🎯 indicates perfect alignment with the evaluation goal.
Each metric includes a 💬 Reason to explain the score — showing what the AI did well ✅ or where it failed ❌.

1️⃣ 💬 Answer Relevancy

• 🎯 Purpose: Measures how relevant the AI-generated answer is to the original input prompt.
• 📊 Scoring: 0.0 – 1.0
• 🧠 Reason: Explains whether the output directly addresses the user query.
  🔻 Low scores = off-topic or partially relevant answers.

2️⃣ 📚 Faithfulness

• 🎯 Purpose: Assesses whether the AI output is factually correct ✅ relative to the source/reference.
• 📊 Scoring: 0.0 – 1.0
• 🔍 Reason: Highlights factual errors ❌ or unsupported claims.
  💯 High scores = answer faithfully reflects expected content.

3️⃣ 🎯 Contextual Precision

• 🎯 Purpose: Measures the proportion of relevant content in the AI output compared to all content generated.
• 📊 Scoring: 0.0 – 1.0
• 🔎 Reason: Indicates if irrelevant or extra content was produced.
  ⚠️ Low precision = unnecessary or inaccurate info included.

4️⃣ 📈 Contextual Recall

• 🎯 Purpose: Measures how much of the expected information is actually present in the AI output.
• 📊 Scoring: 0.0 – 1.0
• 🧾 Reason: Explains which key points were missing.
  🚀 High recall = most expected content covered.

5️⃣ 🔗 Contextual Relevancy

• 🎯 Purpose: Combines Precision 🎯 and Recall 📈 to evaluate how relevant and complete the content is in context.
• 📊 Scoring: 0.0 – 1.0
• 🧩 Reason: Highlights missing or irrelevant pieces, giving a holistic view 🌍 of content quality.

6️⃣ 🚫 Hallucination

• 🎯 Purpose: Detects whether the AI produced fabricated 🌀 or unsupported content.
• 📊 Scoring: 0.0 – 1.0
• 🔍 Reason: High score ✅ = no hallucinations;
  Low score ❌ = identifies speculative or false content.

7️⃣ 📝 Summarization

• 🎯 Purpose: Evaluates how well the AI output summarizes 📚 the provided context or expected answer.
• 📊 Scoring: 0.0 – 1.0
• 🧠 Reason: Indicates whether key points were retained ✨ or lost, and if the summary is concise and coherent 💬.