""" LangChain + RAGAS Integrated App Main application using LangChain for models and RAGAS for evaluation. """ import gradio as gr import pandas as pd import os from typing import List, Tuple, Optional from langchain_evaluator import langchain_evaluator from langchain_models import langchain_models_registry def get_available_datasets() -> List[str]: """Get list of available datasets.""" datasets = [] for item in os.listdir("tasks"): if os.path.isdir(f"tasks/{item}") and not item.startswith("."): datasets.append(item) return sorted(datasets) def get_available_dialects() -> List[str]: """Get list of available SQL dialects.""" return ["presto", "bigquery", "snowflake"] def get_available_models() -> List[str]: """Get list of available models.""" return langchain_models_registry.get_available_models() def get_cases_for_dataset(dataset_name: str) -> List[str]: """Get list of cases for a dataset.""" if not dataset_name: return [] try: dataset = langchain_evaluator.load_dataset(dataset_name) cases = [] for case in dataset['cases']: cases.append(f"{case['id']}: {case['question'][:50]}...") return cases except Exception as e: print(f"Error loading cases for {dataset_name}: {e}") return [] def update_case_dropdown(dataset_name: str): """Update case dropdown with new choices and reset value.""" if not dataset_name: return gr.Dropdown(choices=[], value=None) try: dataset = langchain_evaluator.load_dataset(dataset_name) cases = [] for case in dataset['cases']: cases.append(f"{case['id']}: {case['question'][:50]}...") # Return updated dropdown with new choices and no value return gr.Dropdown(choices=cases, value=None) except Exception as e: print(f"Error loading cases for {dataset_name}: {e}") return gr.Dropdown(choices=[], value=None) def run_evaluation( dataset_name: str, dialect: str, case_selection: str, selected_models: List[str] ) -> Tuple[str, pd.DataFrame, dict, str, str, str]: """Run evaluation for selected models on a case.""" print(f"🔍 DEBUG - case_selection type: {type(case_selection)}, value: {case_selection}") print(f"🔍 DEBUG - dataset_name: {dataset_name}, dialect: {dialect}, selected_models: {selected_models}") if not all([dataset_name, dialect, case_selection, selected_models]): return "Please select all required options.", pd.DataFrame(), {}, "" try: # Handle case_selection if it's a list (shouldn't happen but just in case) if isinstance(case_selection, list): print(f"⚠️ WARNING: case_selection is a list, taking first element") case_selection = case_selection[0] if case_selection else "" # Extract case ID from selection case_id = case_selection.split(":")[0] if ":" in case_selection else case_selection print(f"🚀 Starting evaluation:") print(f" Dataset: {dataset_name}") print(f" Dialect: {dialect}") print(f" Case: {case_id}") print(f" Models: {', '.join(selected_models)}") # Run evaluation results = langchain_evaluator.evaluate_models( dataset_name=dataset_name, dialect=dialect, case_id=case_id, model_names=selected_models ) if not results: return "No results generated. Check console for errors.", pd.DataFrame(), {}, "" # Update leaderboard langchain_evaluator.update_leaderboard(results) # Prepare results for display results_data = [] for result in results: results_data.append({ 'Model': result.model_name, 'Reference SQL (Human)': result.reference_sql[:80] + "..." if len(result.reference_sql) > 80 else result.reference_sql, 'Generated SQL (LLM)': result.generated_sql[:80] + "..." if len(result.generated_sql) > 80 else result.generated_sql, 'Composite Score': f"{result.composite_score:.3f}", 'Correctness': f"{result.correctness_exact:.3f}", 'Result Match F1': f"{result.result_match_f1:.3f}", 'Exec Success': f"{result.exec_success:.3f}", 'Latency (ms)': f"{result.latency_ms:.1f}", 'SQL Quality': f"{result.sql_quality:.3f}", 'Semantic Similarity': f"{result.semantic_similarity:.3f}" }) results_df = pd.DataFrame(results_data) # Detailed results detailed_results = {} for result in results: detailed_results[result.model_name] = { 'reference_sql_human': result.reference_sql, 'raw_sql_llm': result.raw_sql, 'cleaned_sql_llm': result.generated_sql, 'question': result.question, 'all_metrics': { 'correctness_exact': result.correctness_exact, 'result_match_f1': result.result_match_f1, 'exec_success': result.exec_success, 'latency_ms': result.latency_ms, 'readability': result.readability, 'dialect_ok': result.dialect_ok, 'sql_quality': result.sql_quality, 'semantic_similarity': result.semantic_similarity, 'structural_similarity': result.structural_similarity, 'composite_score': result.composite_score } } status = f"✅ Evaluation completed! {len(results)} models evaluated." # Get SQL for display (use first result as example) reference_sql = results[0].reference_sql if results else "" generated_sql = results[0].generated_sql if results else "" return status, results_df, detailed_results, "", reference_sql, generated_sql except Exception as e: error_msg = f"❌ Error during evaluation: {str(e)}" print(error_msg) return error_msg, pd.DataFrame(), {}, "", "", "" def get_leaderboard_display() -> pd.DataFrame: """Get leaderboard data for display.""" try: summary = langchain_evaluator.get_leaderboard_summary(top_n=50) if summary.empty: return pd.DataFrame({ 'Rank': ['-'], 'Model': ['No data available'], 'Avg Composite Score': ['-'], 'Avg Correctness': ['-'], 'Avg Result Match F1': ['-'], 'Avg Exec Success': ['-'], 'Avg Latency (ms)': ['-'], 'Avg SQL Quality': ['-'], 'Avg Semantic Similarity': ['-'], 'Avg Structural Similarity': ['-'], 'Cases Evaluated': ['-'] }) # Sort by composite score (highest first) and add ranking summary_sorted = summary.sort_values('composite_score_mean', ascending=False) # Format for display display_data = [] for rank, (model_name, row) in enumerate(summary_sorted.iterrows(), 1): display_row = { 'Rank': rank, 'Model': model_name, 'Avg Composite Score': f"{row['composite_score_mean']:.3f}", 'Avg Correctness': f"{row['correctness_exact_mean']:.3f}", 'Avg Result Match F1': f"{row['result_match_f1_mean']:.3f}", 'Avg Exec Success': f"{row['exec_success_mean']:.3f}", 'Avg Latency (ms)': f"{row['latency_ms_mean']:.1f}", 'Cases Evaluated': int(row['composite_score_count']) } # Add custom metrics columns if they exist if 'sql_quality_mean' in row: display_row['Avg SQL Quality'] = f"{row['sql_quality_mean']:.3f}" if 'semantic_similarity_mean' in row: display_row['Avg Semantic Similarity'] = f"{row['semantic_similarity_mean']:.3f}" if 'structural_similarity_mean' in row: display_row['Avg Structural Similarity'] = f"{row['structural_similarity_mean']:.3f}" display_data.append(display_row) return pd.DataFrame(display_data) except Exception as e: print(f"Error loading leaderboard: {e}") return pd.DataFrame({ 'Rank': ['-'], 'Model': ['Error loading data'], 'Avg Composite Score': ['-'], 'Avg Correctness': ['-'], 'Avg Result Match F1': ['-'], 'Avg Exec Success': ['-'], 'Avg Latency (ms)': ['-'], 'Avg SQL Quality': ['-'], 'Avg Semantic Similarity': ['-'], 'Avg Structural Similarity': ['-'], 'Cases Evaluated': ['-'] }) def run_comprehensive_evaluation( dataset_name: str, dialect: str, selected_models: List[str], max_cases: int ) -> tuple[str, pd.DataFrame, dict, str, str]: """Run comprehensive evaluation across multiple cases.""" if not all([dataset_name, dialect, selected_models]): return "Please select dataset, dialect, and models.", pd.DataFrame(), {}, "", "" try: print(f"🚀 Starting comprehensive evaluation:") print(f" Dataset: {dataset_name}") print(f" Dialect: {dialect}") print(f" Models: {', '.join(selected_models)}") print(f" Max Cases: {max_cases}") results = langchain_evaluator.run_comprehensive_evaluation( dataset_name=dataset_name, dialect=dialect, model_names=selected_models, max_cases=max_cases if max_cases > 0 else None ) # Update leaderboard langchain_evaluator.update_leaderboard(results) # Prepare results for display results_data = [] for result in results: results_data.append({ 'Model': result.model_name, 'Case': result.case_id, 'Reference SQL (Human)': result.reference_sql[:80] + "..." if len(result.reference_sql) > 80 else result.reference_sql, 'Generated SQL (LLM)': result.generated_sql[:80] + "..." if len(result.generated_sql) > 80 else result.generated_sql, 'Composite Score': f"{result.composite_score:.3f}", 'Correctness': f"{result.correctness_exact:.3f}", 'Result Match F1': f"{result.result_match_f1:.3f}", 'Exec Success': f"{result.exec_success:.3f}", 'Latency (ms)': f"{result.latency_ms:.1f}", 'SQL Quality': f"{result.sql_quality:.3f}", 'Semantic Similarity': f"{result.semantic_similarity:.3f}" }) results_df = pd.DataFrame(results_data) # Detailed results detailed_results = {} for result in results: detailed_results[f"{result.model_name}_{result.case_id}"] = { 'reference_sql_human': result.reference_sql, 'raw_sql_llm': result.raw_sql, 'cleaned_sql_llm': result.generated_sql, 'question': result.question, 'all_metrics': { 'correctness_exact': result.correctness_exact, 'result_match_f1': result.result_match_f1, 'exec_success': result.exec_success, 'latency_ms': result.latency_ms, 'readability': result.readability, 'dialect_ok': result.dialect_ok, 'sql_quality': result.sql_quality, 'semantic_similarity': result.semantic_similarity, 'structural_similarity': result.structural_similarity, 'composite_score': result.composite_score } } status_msg = f"✅ Comprehensive evaluation completed! {len(results)} evaluations performed." # Get SQL for display (use first result as example) reference_sql = results[0].reference_sql if results else "" generated_sql = results[0].generated_sql if results else "" return status_msg, results_df, detailed_results, reference_sql, generated_sql except Exception as e: error_msg = f"❌ Error during comprehensive evaluation: {str(e)}" print(error_msg) return error_msg, pd.DataFrame(), {}, "", "" def create_interface(): """Create the Gradio interface.""" with gr.Blocks(title="NL→SQL Leaderboard (LangChain + RAGAS)", theme=gr.themes.Soft()) as app: gr.Markdown(""" # NL→SQL Leaderboard (LangChain + RAGAS) A comprehensive evaluation platform for English → SQL tasks using LangChain for model management and RAGAS for advanced evaluation metrics. Select a dataset, dialect, and test case, then choose models to evaluate. Results are automatically added to the public leaderboard with RAGAS metrics. """) with gr.Row(): with gr.Column(scale=10): pass # Empty column for spacing with gr.Column(scale=1): refresh_button = gr.Button("Refresh Leaderboard", variant="secondary", size="sm") with gr.Tabs(): # Info Tab (moved to first) with gr.Tab("Info"): gr.Markdown(""" ## About the NL→SQL Leaderboard (LangChain + Custom Evaluation) This platform evaluates natural language to SQL generation using advanced tools: **Technology Stack:** - **LangChain**: Model management and prompt handling - **Custom Evaluation**: Comprehensive evaluation metrics without external dependencies - **Gradio**: User interface - **DuckDB**: SQL execution - **sqlglot**: SQL dialect transpilation - **HuggingFace Transformers**: Local model inference **Features:** - **Local-first approach**: All models run locally for privacy and reliability - **Advanced metrics**: Custom SQL quality, semantic similarity, structural analysis - **Comprehensive evaluation**: Batch processing across multiple cases - **Multi-dialect support**: Presto, BigQuery, and Snowflake SQL dialects - **Real-time leaderboard**: Track model performance across different datasets **Evaluation Metrics:** - **Correctness**: Exact match with reference SQL - **Result Match F1**: Semantic similarity of query results - **Execution Success**: Whether the generated SQL executes without errors - **SQL Quality**: Structural and syntactic quality assessment - **Semantic Similarity**: Meaning-based comparison with reference - **Composite Score**: Weighted combination of all metrics """) # Evaluation Tab with gr.Tab("Evaluate"): with gr.Row(): with gr.Column(scale=1): dataset_dropdown = gr.Dropdown( choices=get_available_datasets(), label="Dataset", value=None, allow_custom_value=True ) dialect_dropdown = gr.Dropdown( choices=get_available_dialects(), label="SQL Dialect", value="presto" ) case_dropdown = gr.Dropdown( choices=[], label="Test Case", interactive=True, value=None, allow_custom_value=False, multiselect=False, info="Select a dataset first to load test cases" ) models_checkbox = gr.CheckboxGroup( choices=get_available_models(), label="Models to Evaluate", value=[] ) run_button = gr.Button("Run Evaluation", variant="primary") with gr.Column(scale=2): status_output = gr.Textbox(label="Status", interactive=False) results_table = gr.Dataframe(label="Run Results", interactive=False) detailed_results = gr.JSON(label="Detailed Metrics", visible=False) # SQL Display Section with gr.Row(): with gr.Column(): reference_sql_display = gr.Code( label="Reference SQL (Human)", language="sql", interactive=False, visible=False ) with gr.Column(): generated_sql_display = gr.Code( label="Generated SQL (LLM)", language="sql", interactive=False, visible=False ) # Metric Explanations with gr.Accordion("📊 How Metrics Are Calculated", open=False): gr.Markdown(""" ### Evaluation Metrics Explained **🎯 Composite Score (0.0 - 1.0)** - Weighted combination of all metrics: `Correctness × 0.3 + Result Match F1 × 0.3 + Exec Success × 0.2 + SQL Quality × 0.1 + Semantic Similarity × 0.1` - Higher is better (1.0 = perfect) **✅ Correctness (0.0 - 1.0)** - Exact string match between generated SQL and reference SQL - 1.0 = identical, 0.0 = completely different **📊 Result Match F1 (0.0 - 1.0)** - F1 score comparing query results (not SQL text) - Executes both SQLs and compares result sets - 1.0 = identical results, 0.0 = completely different results **⚡ Exec Success (0.0 - 1.0)** - Whether the generated SQL executes without errors - 1.0 = executes successfully, 0.0 = execution fails **⏱️ Latency (milliseconds)** - Time taken to generate and execute the SQL - Lower is better (faster response) **🔍 SQL Quality (0.0 - 1.0)** - How well the SQL addresses the question - Based on semantic analysis of question vs SQL intent **🧠 Semantic Similarity (0.0 - 1.0)** - Semantic similarity between generated and reference SQL - Uses sentence transformers to compare meaning - 1.0 = identical meaning, 0.0 = completely different meaning """) # Event handlers dataset_dropdown.change( fn=update_case_dropdown, inputs=[dataset_dropdown], outputs=[case_dropdown] ) run_button.click( fn=run_evaluation, inputs=[dataset_dropdown, dialect_dropdown, case_dropdown, models_checkbox], outputs=[status_output, results_table, detailed_results, gr.State(), reference_sql_display, generated_sql_display] ) # Comprehensive Evaluation Tab with gr.Tab("Comprehensive Evaluation"): with gr.Row(): with gr.Column(scale=1): comp_dataset_dropdown = gr.Dropdown( choices=get_available_datasets(), label="Dataset", value=None, allow_custom_value=True ) comp_dialect_dropdown = gr.Dropdown( choices=get_available_dialects(), label="SQL Dialect", value="presto" ) comp_models_checkbox = gr.CheckboxGroup( choices=get_available_models(), label="Models to Evaluate", value=[] ) max_cases_slider = gr.Slider( minimum=1, maximum=50, value=10, step=1, label="Max Cases to Evaluate" ) comp_run_button = gr.Button("Run Comprehensive Evaluation", variant="primary") with gr.Column(scale=2): comp_status_output = gr.Textbox(label="Status", interactive=False) comp_results_table = gr.Dataframe(label="Comprehensive Results", interactive=False) comp_detailed_results = gr.JSON(label="Detailed Metrics", visible=False) # SQL Display Section for Comprehensive Results with gr.Row(): with gr.Column(): comp_reference_sql_display = gr.Code( label="Reference SQL (Human)", language="sql", interactive=False, visible=False ) with gr.Column(): comp_generated_sql_display = gr.Code( label="Generated SQL (LLM)", language="sql", interactive=False, visible=False ) # Metric Explanations for Comprehensive Evaluation with gr.Accordion("📊 How Metrics Are Calculated", open=False): gr.Markdown(""" ### Comprehensive Evaluation Metrics **🎯 Composite Score (0.0 - 1.0)** - Weighted combination: `Correctness × 0.3 + Result Match F1 × 0.3 + Exec Success × 0.2 + SQL Quality × 0.1 + Semantic Similarity × 0.1` - Higher is better (1.0 = perfect) **✅ Correctness (0.0 - 1.0)** - Exact string match between generated SQL and reference SQL - 1.0 = identical, 0.0 = completely different **📊 Result Match F1 (0.0 - 1.0)** - F1 score comparing query results (not SQL text) - Executes both SQLs and compares result sets - 1.0 = identical results, 0.0 = completely different results **⚡ Exec Success (0.0 - 1.0)** - Whether the generated SQL executes without errors - 1.0 = executes successfully, 0.0 = execution fails **⏱️ Latency (milliseconds)** - Time taken to generate and execute the SQL - Lower is better (faster response) **🔍 SQL Quality (0.0 - 1.0)** - How well the SQL addresses the question - Based on semantic analysis of question vs SQL intent **🧠 Semantic Similarity (0.0 - 1.0)** - Semantic similarity between generated and reference SQL - Uses sentence transformers to compare meaning - 1.0 = identical meaning, 0.0 = completely different meaning **📈 Comprehensive Evaluation** - Tests models across multiple cases and datasets - Provides average performance metrics - Shows consistency across different SQL complexity levels """) comp_run_button.click( fn=run_comprehensive_evaluation, inputs=[comp_dataset_dropdown, comp_dialect_dropdown, comp_models_checkbox, max_cases_slider], outputs=[comp_status_output, comp_results_table, comp_detailed_results, comp_reference_sql_display, comp_generated_sql_display] ) # Leaderboard Tab with gr.Tab("Leaderboard"): leaderboard_table = gr.Dataframe( label="Global Leaderboard (Top 50)", interactive=False, value=get_leaderboard_display() ) # Metric Explanations for Leaderboard with gr.Accordion("📊 How Leaderboard Metrics Are Calculated", open=False): gr.Markdown(""" ### Global Leaderboard Metrics **🏆 Rank** - Models ranked by average composite score (highest first) - Based on aggregated performance across all evaluations **🎯 Avg Composite Score (0.0 - 1.0)** - Average of all composite scores for each model - Weighted combination: `Correctness × 0.3 + Result Match F1 × 0.3 + Exec Success × 0.2 + SQL Quality × 0.1 + Semantic Similarity × 0.1` - Higher is better (1.0 = perfect) **✅ Avg Correctness (0.0 - 1.0)** - Average exact string match between generated SQL and reference SQL - 1.0 = identical, 0.0 = completely different **📊 Avg Result Match F1 (0.0 - 1.0)** - Average F1 score comparing query results (not SQL text) - Executes both SQLs and compares result sets - 1.0 = identical results, 0.0 = completely different results **⚡ Avg Exec Success (0.0 - 1.0)** - Average success rate of SQL execution - 1.0 = always executes successfully, 0.0 = always fails **⏱️ Avg Latency (milliseconds)** - Average time taken to generate and execute SQL - Lower is better (faster response) **📈 Cases Evaluated** - Number of test cases each model has been evaluated on - More cases = more reliable performance metrics **🔍 Avg SQL Quality (0.0 - 1.0)** - Average quality score of how well SQL addresses questions - Based on semantic analysis of question vs SQL intent **🧠 Avg Semantic Similarity (0.0 - 1.0)** - Average semantic similarity between generated and reference SQL - Uses sentence transformers to compare meaning - 1.0 = identical meaning, 0.0 = completely different meaning **📊 Avg Structural Similarity (0.0 - 1.0)** - Average structural similarity between generated and reference SQL - Compares SQL structure, keywords, and patterns - 1.0 = identical structure, 0.0 = completely different structure """) # Add refresh button click event refresh_button.click( fn=get_leaderboard_display, outputs=[leaderboard_table] ) return app if __name__ == "__main__": app = create_interface() app.launch(server_name="0.0.0.0", server_port=7860, share=True)