axelsirota's picture
Upload folder using huggingface_hub
e0f1828 verified
"""
Evaluation Suite Builder - Design custom evaluation frameworks for AI features
Helps PMs create evaluation rubrics, test cases, and judge prompts
"""
import gradio as gr
import json
# Use case templates with recommended evaluation approaches
USE_CASE_TEMPLATES = {
"customer_service": {
"name": "Customer Service Bot",
"description": "AI that responds to customer inquiries and resolves issues",
"priority_dimensions": ["helpfulness", "accuracy", "tone", "resolution"],
"recommended_metrics": [
"Task completion rate",
"Customer satisfaction (CSAT)",
"First-response resolution rate",
"Escalation rate",
"Response relevance score"
],
"test_case_categories": [
"Simple FAQ questions",
"Order status inquiries",
"Complaint handling",
"Refund requests",
"Technical troubleshooting",
"Edge cases (angry customers, unusual requests)"
],
"judge_focus": "Focus on whether the response would actually resolve the customer's issue. Tone should be professional and empathetic. Accuracy of information is critical."
},
"content_generation": {
"name": "Content Generation",
"description": "AI that creates marketing copy, blog posts, or social media content",
"priority_dimensions": ["creativity", "brand_voice", "engagement", "accuracy"],
"recommended_metrics": [
"Brand voice consistency score",
"Engagement prediction",
"Factual accuracy",
"Readability score",
"Human edit rate"
],
"test_case_categories": [
"Product descriptions",
"Social media posts",
"Email subject lines",
"Blog introductions",
"Ad copy variations",
"Edge cases (sensitive topics, competitor mentions)"
],
"judge_focus": "Evaluate creativity and engagement potential while maintaining brand voice. Check for factual claims that could be problematic."
},
"code_assistant": {
"name": "Code Assistant",
"description": "AI that helps developers write, debug, and explain code",
"priority_dimensions": ["correctness", "efficiency", "clarity", "security"],
"recommended_metrics": [
"Code execution success rate",
"Bug introduction rate",
"Code review pass rate",
"Explanation clarity score",
"Security vulnerability detection"
],
"test_case_categories": [
"Simple function implementation",
"Bug fixing",
"Code explanation",
"Refactoring suggestions",
"Security-sensitive code",
"Edge cases (ambiguous requirements, legacy code)"
],
"judge_focus": "Code must be correct and runnable. Security issues are critical failures. Explanations should be clear to developers of varying skill levels."
},
"analysis_assistant": {
"name": "Analysis Assistant",
"description": "AI that analyzes data, documents, or situations and provides insights",
"priority_dimensions": ["accuracy", "completeness", "clarity", "actionability"],
"recommended_metrics": [
"Factual accuracy rate",
"Key insight coverage",
"Actionable recommendation rate",
"Source citation accuracy",
"Logical consistency score"
],
"test_case_categories": [
"Data summarization",
"Trend identification",
"Comparison analysis",
"Risk assessment",
"Recommendation generation",
"Edge cases (insufficient data, conflicting information)"
],
"judge_focus": "Analysis must be logically sound and factually accurate. Conclusions should be supported by the data. Recommendations should be specific and actionable."
},
"rag_knowledge_base": {
"name": "RAG Knowledge Base",
"description": "AI that answers questions from your company's documents (knowledge base, help center, internal docs)",
"priority_dimensions": ["accuracy", "relevance", "completeness", "groundedness"],
"recommended_metrics": [
"Faithfulness score (are claims supported by retrieved docs?)",
"Answer relevancy (does it answer the actual question?)",
"Context precision (did we retrieve the right docs?)",
"Context recall (did we miss relevant docs?)",
"Hallucination rate (claims not in source material)"
],
"test_case_categories": [
"Simple fact lookup (who, what, when)",
"Multi-document synthesis",
"Questions with no good answer in docs",
"Ambiguous questions",
"Questions requiring recent information",
"Edge cases (jargon, abbreviations, typos)"
],
"judge_focus": "Focus on whether claims are supported by the retrieved documents. Hallucination is the critical failure mode. Check that the answer directly addresses the question asked, not a related question."
},
"ai_agent": {
"name": "AI Agent / Automation",
"description": "AI that takes actions autonomously (booking, scheduling, data entry, multi-step workflows)",
"priority_dimensions": ["accuracy", "safety", "completeness", "task_completion"],
"recommended_metrics": [
"Task completion rate (did it finish the job?)",
"Goal accuracy (did it solve the RIGHT problem?)",
"Steps to completion (efficiency)",
"Escalation rate (how often does it need human help?)",
"Error recovery rate (can it fix its own mistakes?)",
"Tool selection accuracy (right tool for the job?)"
],
"test_case_categories": [
"Happy path (standard requests)",
"Multi-step workflows",
"Ambiguous instructions",
"Requests requiring clarification",
"Requests outside scope",
"Error scenarios (API failures, missing data)",
"Adversarial inputs (prompt injection attempts)"
],
"judge_focus": "Grade the OUTCOME, not the path. Did the agent accomplish what the user wanted? Did it ask for clarification when needed? Did it avoid taking irreversible actions without confirmation?"
}
}
# Evaluation dimension definitions
DIMENSION_DEFINITIONS = {
"accuracy": {
"name": "Accuracy",
"description": "Is the information factually correct and verifiable?",
"rubric_1": "Contains significant factual errors",
"rubric_3": "Mostly accurate with minor errors",
"rubric_5": "Completely accurate and verifiable"
},
"relevance": {
"name": "Relevance",
"description": "Does the response address what was actually asked?",
"rubric_1": "Completely off-topic",
"rubric_3": "Partially addresses the question",
"rubric_5": "Directly and fully addresses the question"
},
"helpfulness": {
"name": "Helpfulness",
"description": "Would this actually help the user accomplish their goal?",
"rubric_1": "Not helpful at all",
"rubric_3": "Somewhat helpful but incomplete",
"rubric_5": "Extremely helpful and actionable"
},
"clarity": {
"name": "Clarity",
"description": "Is the response well-organized and easy to understand?",
"rubric_1": "Confusing and poorly structured",
"rubric_3": "Understandable with some effort",
"rubric_5": "Crystal clear and well-organized"
},
"tone": {
"name": "Tone",
"description": "Is the tone appropriate for the context?",
"rubric_1": "Inappropriate tone",
"rubric_3": "Acceptable but could be better",
"rubric_5": "Perfect tone for the situation"
},
"completeness": {
"name": "Completeness",
"description": "Does the response cover all important aspects?",
"rubric_1": "Missing critical information",
"rubric_3": "Covers main points but missing some details",
"rubric_5": "Comprehensive and thorough"
},
"creativity": {
"name": "Creativity",
"description": "Is the response original and engaging?",
"rubric_1": "Generic and uninspired",
"rubric_3": "Shows some creativity",
"rubric_5": "Highly creative and engaging"
},
"safety": {
"name": "Safety",
"description": "Is the response free from harmful or inappropriate content?",
"rubric_1": "Contains harmful content",
"rubric_3": "Some potentially concerning elements",
"rubric_5": "Completely safe and appropriate"
},
"groundedness": {
"name": "Groundedness",
"description": "Is every claim in the response supported by the retrieved context?",
"rubric_1": "Multiple unsupported claims (hallucination)",
"rubric_3": "Mostly grounded with minor unsupported details",
"rubric_5": "Every claim is traceable to source documents"
},
"task_completion": {
"name": "Task Completion",
"description": "Did the agent successfully complete the requested task?",
"rubric_1": "Task failed or abandoned",
"rubric_3": "Partially completed or required human intervention",
"rubric_5": "Task fully completed as requested"
}
}
def load_template(use_case):
"""Load a use case template"""
if use_case not in USE_CASE_TEMPLATES:
return "", "", [], "", ""
template = USE_CASE_TEMPLATES[use_case]
return (
template["name"],
template["description"],
template["priority_dimensions"],
"\n".join(f"- {m}" for m in template["recommended_metrics"]),
template["judge_focus"]
)
def generate_evaluation_suite(
feature_name, feature_description, selected_dimensions,
sample_size, evaluator_type, include_test_cases
):
"""Generate a complete evaluation suite"""
if not feature_name or not selected_dimensions:
return "Please provide a feature name and select at least one evaluation dimension."
# Build the evaluation suite
suite = f"# Evaluation Suite: {feature_name}\n\n"
suite += f"**Feature Description:** {feature_description}\n\n"
# Evaluation dimensions section
suite += "## Evaluation Dimensions\n\n"
for dim in selected_dimensions:
if dim in DIMENSION_DEFINITIONS:
d = DIMENSION_DEFINITIONS[dim]
suite += f"### {d['name']}\n"
suite += f"{d['description']}\n\n"
suite += "**Rubric:**\n"
suite += f"- 1-2: {d['rubric_1']}\n"
suite += f"- 3-4: {d['rubric_3']}\n"
suite += f"- 5: {d['rubric_5']}\n\n"
# Sample size and evaluator recommendations
suite += "## Evaluation Setup\n\n"
suite += f"**Recommended Sample Size:** {sample_size} examples\n"
suite += f"**Evaluator Type:** {evaluator_type}\n\n"
if evaluator_type == "Human Only":
suite += "**Evaluator Requirements:**\n"
suite += "- 2-3 evaluators per sample for inter-rater reliability\n"
suite += "- Calibration session before evaluation begins\n"
suite += "- 10-20% overlap for reliability measurement\n"
suite += "- Target Fleiss' Kappa > 0.4\n\n"
elif evaluator_type == "LLM-as-Judge Only":
suite += "**LLM Judge Setup:**\n"
suite += "- Use GPT-4 or Claude as judge model\n"
suite += "- Validate against human ratings on 50+ samples first\n"
suite += "- Monitor for position and length bias\n"
suite += "- Re-calibrate monthly\n\n"
else:
suite += "**Hybrid Approach:**\n"
suite += "- LLM judge for initial screening (100% of samples)\n"
suite += "- Human review for flagged samples and random 5-10%\n"
suite += "- Calibrate LLM judge quarterly against human ratings\n\n"
# Test cases section
if include_test_cases:
suite += "## Test Case Categories\n\n"
suite += "Create test cases for each category:\n\n"
suite += "1. **Happy Path** - Standard, expected inputs\n"
suite += "2. **Edge Cases** - Unusual but valid inputs\n"
suite += "3. **Adversarial** - Attempts to break or manipulate\n"
suite += "4. **Domain-Specific** - Industry or use-case specific scenarios\n"
suite += "5. **Failure Recovery** - How does it handle errors?\n\n"
# Judge prompt section
suite += "## LLM Judge Prompt Template\n\n"
suite += "```\n"
suite += f"You are evaluating an AI response for a {feature_name}.\n\n"
suite += "**Evaluation Criteria:**\n"
for dim in selected_dimensions:
if dim in DIMENSION_DEFINITIONS:
d = DIMENSION_DEFINITIONS[dim]
suite += f"- {d['name']}: {d['description']}\n"
suite += "\n"
suite += "**Scoring:**\n"
suite += "Rate each criterion from 1-5 with a brief explanation.\n"
suite += "```\n"
return suite
def export_as_json(
feature_name, feature_description, selected_dimensions,
sample_size, evaluator_type
):
"""Export evaluation config as JSON"""
config = {
"feature_name": feature_name,
"feature_description": feature_description,
"dimensions": selected_dimensions,
"sample_size": sample_size,
"evaluator_type": evaluator_type,
"dimension_rubrics": {
dim: DIMENSION_DEFINITIONS[dim]
for dim in selected_dimensions
if dim in DIMENSION_DEFINITIONS
}
}
return json.dumps(config, indent=2)
# Build Gradio interface
with gr.Blocks(title="Evaluation Suite Builder", theme=gr.themes.Soft()) as demo:
gr.Markdown("""
# Evaluation Suite Builder
Design custom evaluation frameworks for your AI features. Get rubrics, sample sizes,
evaluator recommendations, and judge prompts.
**For Product Managers:** Use this to create evaluation plans before launching AI features.
""")
gr.Markdown(
"> **PM Decision:** Your evaluation suite defines what 'good' means for your AI. "
"Build it before you build the system - changing success criteria mid-project is expensive."
)
with gr.Row():
with gr.Column(scale=1):
gr.Markdown("### 1. Feature Details")
template_dropdown = gr.Dropdown(
choices=[
("Customer Service Bot", "customer_service"),
("Content Generation", "content_generation"),
("Code Assistant", "code_assistant"),
("Analysis Assistant", "analysis_assistant"),
("RAG Knowledge Base", "rag_knowledge_base"),
("AI Agent / Automation", "ai_agent")
],
label="Load Template (Optional)",
value=None
)
feature_name = gr.Textbox(
label="Feature Name",
placeholder="e.g., Customer Support Chatbot"
)
feature_description = gr.Textbox(
label="Feature Description",
placeholder="What does this AI feature do?",
lines=3
)
gr.Markdown("### 2. Evaluation Dimensions")
dimensions = gr.CheckboxGroup(
choices=[
("Accuracy", "accuracy"),
("Relevance", "relevance"),
("Helpfulness", "helpfulness"),
("Clarity", "clarity"),
("Tone", "tone"),
("Completeness", "completeness"),
("Creativity", "creativity"),
("Safety", "safety"),
("Groundedness (RAG)", "groundedness"),
("Task Completion (Agents)", "task_completion")
],
value=["accuracy", "relevance", "helpfulness"],
label="Select Dimensions to Evaluate"
)
gr.Markdown("### 3. Evaluation Setup")
sample_size = gr.Radio(
choices=["50 (Quick pilot)", "100 (Standard)", "250 (Thorough)", "500+ (Comprehensive)"],
value="100 (Standard)",
label="Sample Size"
)
evaluator_type = gr.Radio(
choices=["Human Only", "LLM-as-Judge Only", "Hybrid (Recommended)"],
value="Hybrid (Recommended)",
label="Evaluator Type"
)
include_test_cases = gr.Checkbox(
value=True,
label="Include Test Case Categories"
)
generate_btn = gr.Button("Generate Evaluation Suite", variant="primary")
with gr.Column(scale=2):
gr.Markdown("### Your Evaluation Suite")
output = gr.Markdown()
with gr.Accordion("Export as JSON", open=False):
json_output = gr.Code(language="json")
export_btn = gr.Button("Export Config")
gr.Markdown("""
---
### PM Checklist: Before Launching AI
- [ ] Evaluation dimensions defined and prioritized
- [ ] Rubric created with clear scoring criteria
- [ ] Test cases covering happy path, edge cases, and adversarial inputs
- [ ] Evaluator setup (human, LLM, or hybrid)
- [ ] Baseline established from current solution or competitor
- [ ] Success criteria defined (what score = launch-ready?)
- [ ] Monitoring plan for post-launch evaluation
""")
# Event handlers
template_dropdown.change(
fn=load_template,
inputs=[template_dropdown],
outputs=[feature_name, feature_description, dimensions, gr.Textbox(visible=False), gr.Textbox(visible=False)]
)
generate_btn.click(
fn=generate_evaluation_suite,
inputs=[feature_name, feature_description, dimensions, sample_size, evaluator_type, include_test_cases],
outputs=[output]
)
export_btn.click(
fn=export_as_json,
inputs=[feature_name, feature_description, dimensions, sample_size, evaluator_type],
outputs=[json_output]
)
if __name__ == "__main__":
demo.launch()