Reexpress MCP Server

# Copyright Reexpress AI, Inc. All rights reserved. import torch import html import argparse import constants import mcp_utils_test import utils_visualization__constants def create_html_page(current_reexpression, nearest_match_meta_data=None, nearest_match_meta_data_is_html_escaped=False): """ Creates a static HTML page from model output dictionary. Args: current_reexpression: Dictionary containing the model output with keys for each field nearest_match_meta_data_is_html_escaped: True if the text fields have been processed with html.escape() Returns: HTML string """ # Extract verification results # Need defaults for testing prediction_meta_data = current_reexpression.get("prediction_meta_data", {}) predicted_class = prediction_meta_data.get("prediction", 0) successfully_verified = predicted_class == 1 if successfully_verified: successfully_verified_html_class = "positive" else: successfully_verified_html_class = "negative" is_high_reliability_region = prediction_meta_data.get("is_high_reliability_region", False) is_ood = prediction_meta_data.get("is_ood", True) is_ood_html_class = "positive" if not is_ood else "negative" calibration_reliability = \ mcp_utils_test.get_calibration_reliability_label(is_high_reliability_region, is_ood) # Model Level try: hr_class_conditional_accuracy = prediction_meta_data["hr_class_conditional_accuracy"] min_rescaled_similarity_to_determine_high_reliability_region = \ prediction_meta_data["min_rescaled_similarity_to_determine_high_reliability_region"] hr_output_thresholds = \ prediction_meta_data["hr_output_thresholds"] support_index_ntotal = prediction_meta_data["support_index_ntotal"] except: hr_class_conditional_accuracy = 0.0 min_rescaled_similarity_to_determine_high_reliability_region = "N/A" hr_output_thresholds = "N/A" support_index_ntotal = "N/A" classification_confidence, classification_confidence_html_class = \ mcp_utils_test.get_calibration_confidence_label(calibration_reliability=calibration_reliability, hr_class_conditional_accuracy=hr_class_conditional_accuracy, return_html_class=True) model1_name = "gpt-5-2025-08-07" model2_name = "gemini-2.5-pro" agreement_model_name = "granite-3.3-8b-instruct" model1_classification = current_reexpression.get(constants.REEXPRESS_MODEL1_CLASSIFICATION, False) model2_classification = current_reexpression.get(constants.REEXPRESS_MODEL2_CLASSIFICATION, False) agreement_model_classification = \ current_reexpression.get(constants.REEXPRESS_AGREEMENT_MODEL_CLASSIFICATION, False) if agreement_model_classification: agreement_model_classification_string = "Yes" else: agreement_model_classification_string = "No" model1_html_class = "positive" if model1_classification else "negative" model2_html_class = "positive" if model2_classification else "negative" agreement_model_html_class = "positive" if agreement_model_classification else "negative" # We escape HTML as it may be contained within the responses themselves model1_explanation = html.escape(current_reexpression.get(constants.REEXPRESS_MODEL1_EXPLANATION, '')) model2_explanation = html.escape(current_reexpression.get(constants.REEXPRESS_MODEL2_EXPLANATION, '')) model1_summary = html.escape(current_reexpression.get(constants.REEXPRESS_MODEL1_TOPIC_SUMMARY, '')) files_in_consideration_message = \ mcp_utils_test.get_files_in_consideration_message( current_reexpression.get(constants.REEXPRESS_ATTACHED_FILE_NAMES, [])).strip() submitted_time = current_reexpression.get(constants.REEXPRESS_SUBMITTED_TIME_KEY, 'N/A') # Uncertainty try: sdm_output = \ prediction_meta_data["sdm_output"].detach().cpu().tolist() # TODO: Move this earlier to avoid duplication: is_high_reliability_region = prediction_meta_data["is_high_reliability_region"] is_high_reliability_region_html_class = "positive" if is_high_reliability_region else "negative" rescaled_similarity = prediction_meta_data["rescaled_similarity"] cumulative_effective_sample_sizes = \ prediction_meta_data["cumulative_effective_sample_sizes"].detach().cpu().tolist() similarity_q = int(prediction_meta_data["q"]) distance_d = prediction_meta_data["d"] magnitude = prediction_meta_data["f"].detach().cpu().tolist() # analysis of the effective sample size: distance_d_lower = prediction_meta_data["d_lower"] distance_d_upper = prediction_meta_data["d_upper"] sdm_output_d_lower = prediction_meta_data["sdm_output_d_lower"].detach().cpu().tolist() sdm_output_d_upper = prediction_meta_data["sdm_output_d_upper"].detach().cpu().tolist() except: sdm_output = "N/A" is_high_reliability_region = False rescaled_similarity = "N/A" is_high_reliability_region_html_class = "negative" cumulative_effective_sample_sizes = "N/A" similarity_q = "N/A" distance_d = "N/A" magnitude = "N/A" # analysis of the effective sample size: distance_d_lower = "N/A" distance_d_upper = "N/A" sdm_output_d_lower = "N/A" sdm_output_d_upper = "N/A" user_question = html.escape(current_reexpression.get(constants.REEXPRESS_QUESTION_KEY, '')) ai_response = html.escape(current_reexpression.get(constants.REEXPRESS_AI_RESPONSE_KEY, '')) # Nearest Match try: assert nearest_match_meta_data is not None nearest_match_html_string = nearest_match_html(nearest_match_meta_data, model1_name, model2_name, agreement_model_name, content_is_html_escaped=nearest_match_meta_data_is_html_escaped) except: nearest_match_html_string = """<div class="section" style="margin-left: 40px;"> The nearest match is not available. </div>""" html_content_string = f""" <!DOCTYPE html> <html lang="en"> <head> <meta charset="UTF-8"> <meta name="viewport" content="width=device-width, initial-scale=1.0"> <title>Reexpress MCP Server Output</title> {utils_visualization__constants.css_style} </head> <body> <div class="container"> <div class="header"> Reexpress MCP Server Output <span class="mcp-server-version">(v{constants.REEXPRESS_MCP_SERVER_VERSION})</span> </div> <div class="section"> <div class="section-title">Verification Results</div> <div class="field-box" style="margin-bottom: 20px;"> <div class="field-label">Successfully Verified (Prediction)</div> <div class="field-value"> <div class="field-value"><span class="tag tag-{successfully_verified_html_class}">{successfully_verified}</span></div> </div> </div> <div class="field-box" style="margin-bottom: 20px;"> <div class="field-label">Confidence</div> <div class="field-value"><span class="tag tag-{classification_confidence_html_class}">{classification_confidence}</span></div> </div> <div class="explanation-box-{model1_html_class}"> <div class="explanation-title-{model1_html_class}">Model 1 Summary <span class="model-name">({model1_name})</span></div> <div>{model1_summary}</div> </div> <div class="explanation-box-{model1_html_class}"> <div class="explanation-title-{model1_html_class}">Model 1 Explanation <span class="model-name">({model1_name})</span></div> <div>{model1_explanation}</div> </div> <div class="explanation-box-{model2_html_class}"> <div class="explanation-title-{model2_html_class}">Model 2 Explanation <span class="model-name">({model2_name})</span></div> <div>{model2_explanation}</div> </div> <div class="explanation-box-{agreement_model_html_class}"> <div class="explanation-title-{agreement_model_html_class}">Model 3 Agreement <span class="model-name">({agreement_model_name})</span></div> <div>{constants.AGREEMENT_MODEL_USER_FACING_PROMPT}</div> <div><span class="tag tag-{agreement_model_html_class}">{agreement_model_classification_string}</span></div> </div> </div> <div class="separator"></div> <div class="section"> <div class="section-title">Additional Information</div> <div class="field-grid"> <div class="field-box"> <div class="field-label">File Access</div> <div class="field-value">{files_in_consideration_message}</div> </div> <div class="field-box"> <div class="field-label">Date</div> <div class="field-value">{submitted_time}</div> </div> </div> </div> <div class="section"> <div class="section-title">Uncertainty (instance-level) Details</div> <div class="field-box" style="margin-bottom: 20px;"> <div class="field-label">p(y | x)</div> <div class="field-value">{sdm_output}</div> </div> <div class="field-grid"> <div class="field-box"> <div class="field-label">{constants.CALIBRATION_HIGH_RELIABILITY_REGION_LABEL_FULL}</div> <div class="field-value"> <span class="tag tag-{is_high_reliability_region_html_class}">{is_high_reliability_region}</span> </div> </div> <div class="field-box"> <div class="field-label">Out-of-Distribution</div> <div class="field-value"> <span class="tag tag-{is_ood_html_class}">{is_ood}</span> </div> </div> <div class="field-box"> <div class="field-label">Rescaled Similarity (q')</div> <div class="field-value">{rescaled_similarity}</div> </div> </div> <div class="field-grid"> <div class="field-box"> <div class="field-label"> {constants.qFull} </div> <div class="field-value">{similarity_q}</div> </div> <div class="field-box"> <div class="field-label"> {constants.dQuantileFull} </div> <div class="field-value">{distance_d}</div> </div> <div class="field-box"> <div class="field-label"> {constants.fFull} </div> <div class="field-value">{magnitude}</div> </div> </div> </div> <div class="section" style="margin-left: 40px;"> <div class="section-title">Analysis of the Effective Sample Size</div> <div class="field-box" style="margin-bottom: 20px;"> <div class="field-label">p(y | x)_lower</div> <div class="field-value">{sdm_output_d_lower}</div> </div> <div class="field-box" style="margin-bottom: 20px;"> <div class="field-label">p(y | x)_upper</div> <div class="field-value">{sdm_output_d_upper}</div> </div> <div class="field-grid"> <div class="field-box"> <div class="field-label">Effective Sample Size (by class)</div> <div class="field-value">{cumulative_effective_sample_sizes}</div> </div> </div> <div class="field-grid"> <div class="field-box"> <div class="field-label"> {constants.dQuantileLowerFull} </div> <div class="field-value">{distance_d_lower}</div> </div> <div class="field-box"> <div class="field-label"> {constants.dQuantileUpperFull} </div> <div class="field-value">{distance_d_upper}</div> </div> </div> </div> <div class="section"> <div class="section-title">SDM Estimator (Model-level) Details</div> <div class="field-grid"> <div class="field-box"> <div class="field-label"> α </div> <div class="field-value">{hr_class_conditional_accuracy}</div> </div> <div class="field-box"> <div class="field-label"> Minimum Rescaled Similarity (q'_min) </div> <div class="field-value">{min_rescaled_similarity_to_determine_high_reliability_region}</div> </div> <div class="field-box"> <div class="field-label"> Class-wise Output Thresholds (ψ) </div> <div class="field-value">{hr_output_thresholds}</div> </div> <div class="field-box"> <div class="field-label"> Support/training size </div> <div class="field-value">{support_index_ntotal}</div> </div> </div> </div> <div class="section"> <div class="section-title">Prompt</div> <div class="prompt-box">{user_question}</div> </div> <div class="section"> <div class="section-title">AI Response</div> <div class="document-box">{ai_response}</div> </div> <div class="separator"></div> {nearest_match_html_string} <div class="separator"></div> <div class="section"> <div class="section-title">Legend</div> <div class="legend-content"> <p>An ensemble of models 1, 2, and 3 (including the hidden states of model 3) is taken as the input to the SDM estimator that determines the verification classification.</p> <p>The classification is in the {constants.CALIBRATION_HIGH_RELIABILITY_REGION_LABEL_FULL_NON_TITLE} when the rescaled Similarity (q') is at least the minimum rescaled Similarity (q'_min) AND the predictive uncertainty, p(y | x), for the predicted class is at least the corresponding class-wise output threshold (ψ) for the predicted class.</p> <p>The estimates in the section 'Analysis of the Effective Sample Size' are based on the DKW inequality applied to the distance quantiles.</p> <div class="legend-items"> <div class="legend-item"> <span class="legend-label">Class 0:</span> <span class="legend-value">{constants.MCP_SERVER_NOT_VERIFIED_CLASS_LABEL}</span> </div> <div class="legend-item"> <span class="legend-label">Class 1:</span> <span class="legend-value">{constants.MCP_SERVER_VERIFIED_CLASS_LABEL}</span> </div> </div> </div> </div> </div> </body> </html> """ return html_content_string def escape_html(content, content_is_html_escaped=False): if content_is_html_escaped: return content else: return html.escape(content) def nearest_match_html(nearest_match_meta_data, model1_name, model2_name, agreement_model_name, content_is_html_escaped=False): # Use content_is_html_escaped=True when you know the fields have already been escaped with html.escape() (e.g., # when retrieving from a database) predicted_class = nearest_match_meta_data.get("model_train_predicted_label", -1) true_label = nearest_match_meta_data.get("model_train_label", -1) successfully_verified = predicted_class == 1 if successfully_verified: successfully_verified_html_class = "positive" else: successfully_verified_html_class = "negative" if true_label == 1: true_class_string_label = constants.MCP_SERVER_VERIFIED_CLASS_LABEL true_class_html_class = "positive" else: true_class_string_label = constants.MCP_SERVER_NOT_VERIFIED_CLASS_LABEL true_class_html_class = "negative" model1_classification = nearest_match_meta_data.get("model1_classification_int", -1) == 1 model2_classification = nearest_match_meta_data.get("model2_classification_int", -1) == 1 agreement_model_classification = \ nearest_match_meta_data.get("agreement_model_classification_int", -1) == 1 if agreement_model_classification: agreement_model_classification_string = "Yes" else: agreement_model_classification_string = "No" model1_html_class = "positive" if model1_classification else "negative" model2_html_class = "positive" if model2_classification else "negative" agreement_model_html_class = "positive" if agreement_model_classification else "negative" # Escape HTML, as applicable model1_explanation = escape_html(nearest_match_meta_data.get(constants.REEXPRESS_MODEL1_EXPLANATION, ''), content_is_html_escaped=content_is_html_escaped) model2_explanation = escape_html(nearest_match_meta_data.get(constants.REEXPRESS_MODEL2_EXPLANATION, ''), content_is_html_escaped=content_is_html_escaped) model1_summary = escape_html(nearest_match_meta_data.get(constants.REEXPRESS_MODEL1_TOPIC_SUMMARY, ''), content_is_html_escaped=content_is_html_escaped) user_question = escape_html(nearest_match_meta_data.get("user_question", ''), content_is_html_escaped=content_is_html_escaped) ai_response = escape_html(nearest_match_meta_data.get(constants.REEXPRESS_AI_RESPONSE_KEY, ''), content_is_html_escaped=content_is_html_escaped) document_id = escape_html(nearest_match_meta_data.get("document_id", ""), content_is_html_escaped=content_is_html_escaped) document_source = escape_html(nearest_match_meta_data.get("document_source", ""), content_is_html_escaped=content_is_html_escaped) nearest_match_html_string = f""" <div class="nearest-match-box"> <div class="section" style="margin-left: 40px;"> <div class="section-title">Nearest Match in Training</div> <div class="field-grid"> <div class="field-box" style="margin-bottom: 20px;"> <div class="field-label">Successfully Verified (Prediction)</div> <div class="field-value"> <div class="field-value"><span class="tag tag-{successfully_verified_html_class}">{successfully_verified}</span></div> </div> </div> <div class="field-box" style="margin-bottom: 20px;"> <div class="field-label">True Label</div> <div class="field-value"><span class="tag tag-{true_class_html_class}">{true_class_string_label}</span></div> </div> </div> <div class="explanation-box-{model1_html_class}"> <div class="explanation-title-{model1_html_class}">Model 1 Summary <span class="model-name">({model1_name})</span></div> <div>{model1_summary}</div> </div> <div class="explanation-box-{model1_html_class}"> <div class="explanation-title-{model1_html_class}">Model 1 Explanation <span class="model-name">({model1_name})</span></div> <div>{model1_explanation}</div> </div> <div class="explanation-box-{model2_html_class}"> <div class="explanation-title-{model2_html_class}">Model 2 Explanation <span class="model-name">({model2_name})</span></div> <div>{model2_explanation}</div> </div> <div class="explanation-box-{agreement_model_html_class}"> <div class="explanation-title-{agreement_model_html_class}">Model 3 Agreement <span class="model-name">({agreement_model_name})</span></div> <div>{constants.AGREEMENT_MODEL_USER_FACING_PROMPT}</div> <div><span class="tag tag-{agreement_model_html_class}">{agreement_model_classification_string}</span></div> </div> <div class="section"> <div class="section-title">Prompt</div> <div class="prompt-box">{user_question}</div> </div> <div class="section"> <div class="section-title">AI Response</div> <div class="document-box">{ai_response}</div> </div> <div class="field-grid"> <div class="field-box" style="margin-bottom: 20px;"> <div class="field-label">Document ID</div> <div class="field-value">{document_id}</div> </div> <div class="field-box" style="margin-bottom: 20px;"> <div class="field-label">Document Source</div> <div class="field-value">{document_source}</div> </div> </div> </div> </div> """ return nearest_match_html_string def save_html_file(current_reexpression, nearest_match_meta_data=None, filename='reexpress_mcp_server_output.html'): """ Saves the generated HTML to a file. Args: current_reexpression: Dictionary containing the neural model output filename: Name of the output HTML file """ html_content = create_html_page(current_reexpression, nearest_match_meta_data=nearest_match_meta_data) with open(filename, 'w', encoding='utf-8') as f: f.write(html_content) print(f"HTML file saved as: {filename}") def main(): parser = argparse.ArgumentParser(description="-----[VISUALIZE]-----") parser.add_argument("--output_file", default="", help="") options = parser.parse_args() # Generate and save the HTML file save_html_file({}, {}, filename=options.output_file) if __name__ == "__main__": main()