import pandas as pd import gradio as gr import os import io import json import gspread from huggingface_hub import HfApi, hf_hub_download import torch import torch.nn.functional as F from transformers import AutoTokenizer, AutoModelForSequenceClassification import re from captum.attr import LayerIntegratedGradients, TokenReferenceBase from captum.attr import visualization as viz from huggingface_hub import InferenceClient from datetime import datetime import uuid # HF_TOKEN = os.environ.get("HF_TOKEN", f"{HF}_{token}") HF_TOKEN = os.environ.get("HF_TOKEN") HF_DATASET_REPO = "akaburia/policy-evaluations" HF_CSV_FILE = "policy_coherence_annotations.csv" HF_USERS_FILE = "user_profiles.csv" HF_CHAT_LOG_FILE = "chatbot_logs.csv" # IMPORT GOOGLE CLOUD TRANSLATE try: from google.cloud import translate_v2 as translate except ImportError: raise ImportError("Please install the translation library by running: pip install google-cloud-translate") try: from zoneinfo import ZoneInfo except ImportError: import pytz # Fallback if zoneinfo is missing # --- COMPREHENSIVE LOGGING --- LOG_FILE = "logs.txt" def write_log(action_type, details): """Appends a timestamped log entry to logs.txt""" try: try: tz = ZoneInfo("Africa/Nairobi") except: import pytz tz = pytz.timezone("Africa/Nairobi") timestamp = datetime.now(tz).strftime('%Y-%m-%d %H:%M:%S') log_entry = f"[{timestamp}] [{action_type}] {details}\n" with open(LOG_FILE, "a", encoding="utf-8") as f: f.write(log_entry) print(log_entry.strip()) # Also print to console for debugging except Exception as e: print(f"Logging failed: {e}") # --- CACHING HELPERS --- DRAFT_FILE = "user_drafts.json" def load_drafts(): if os.path.exists(DRAFT_FILE): try: with open(DRAFT_FILE, 'r') as f: return json.load(f) except: return {} return {} def update_cache_row(user, session_id, dom_a, pol_a, dom_b, pol_b, tar_col, ctx_col, a_list, idx, b_text, rel, inter, just): """Fires automatically on keystrokes/clicks to save progress and workspace state""" if not user or not a_list or idx >= len(a_list) or not b_text: return curr_a = a_list[idx] drafts = load_drafts() # Upgraded structure to hold workspace settings AND row data if user not in drafts: drafts[user] = {"workspace": {}, "rows": {}} # Save the active workspace so we can restore it on reload drafts[user]["workspace"] = { "session_id": session_id, "dom_a": dom_a, "pol_a": pol_a, "dom_b": dom_b, "pol_b": pol_b, "tar_col": tar_col, "ctx_col": ctx_col } cache_key = f"{pol_a}|{pol_b}|{curr_a}" if cache_key not in drafts[user]["rows"]: drafts[user]["rows"][cache_key] = {} # Store the exact state of this specific row with the unique session tag drafts[user]["rows"][cache_key][b_text] = { "rel": rel, "inter": inter, "just": just, "session_id": session_id } write_log("CACHE_UPDATE", f"User {user} auto-saved draft for row index {idx}.") with open(DRAFT_FILE, 'w') as f: json.dump(drafts, f) # ========================================== # 0. MODEL PRELOADING & INFERENCE MATH # ========================================== print("Loading Inference Model into Memory...") device = torch.device("cuda" if torch.cuda.is_available() else "cpu") tokenizer = AutoTokenizer.from_pretrained("akaburia/policy-evaluations") model = AutoModelForSequenceClassification.from_pretrained("akaburia/policy-evaluations").to(device) id_to_label = {0: "neutral", 1: "coherent", 2: "incoherent"} def custom_forward(input_ids, attention_mask): inputs_embeds = model.roberta.embeddings.word_embeddings(input_ids) return model(inputs_embeds=inputs_embeds, attention_mask=attention_mask).logits # Explainability lig = LayerIntegratedGradients(custom_forward, model.roberta.embeddings.word_embeddings) llm_client = InferenceClient("Qwen/Qwen3-8B", token=HF_TOKEN) def generate_row_explanation(a_list, idx, text_b, lang): if not a_list or idx >= len(a_list) or not text_b: return "", "", "", "" policy_a = clean_policy_text(a_list[idx]) policy_b = clean_policy_text(text_b) # 1. Run Captum Explainer model.zero_grad() inputs = tokenizer(policy_a, policy_b, return_tensors="pt", truncation=True, max_length=256) input_ids = inputs["input_ids"].to(device) attention_mask = inputs["attention_mask"].to(device) ref_token_id = tokenizer.pad_token_id special_token_mask = [1 if id in tokenizer.all_special_ids else 0 for id in input_ids[0].tolist()] baseline_ids = torch.tensor([[id if is_special else ref_token_id for id, is_special in zip(input_ids[0].tolist(), special_token_mask)]]).to(device) with torch.no_grad(): logits = model(input_ids=input_ids, attention_mask=attention_mask).logits predicted_class_idx = torch.argmax(logits, dim=1).item() prediction = id_to_label[predicted_class_idx] attributions, _ = lig.attribute( inputs=input_ids, baselines=baseline_ids, additional_forward_args=(attention_mask,), target=predicted_class_idx, return_convergence_delta=True ) attributions = attributions.sum(dim=-1).squeeze(0) attributions = attributions / torch.norm(attributions) attributions = attributions.cpu().detach().numpy() tokens = tokenizer.convert_ids_to_tokens(input_ids[0]) ig_dict = {t.replace('Ġ', '').strip(): float(s) for t, s in zip(tokens, attributions) if t.replace('Ġ', '').strip()} ig_json_str = json.dumps(ig_dict) score_list = [f"'{k}': {v:.3f}" for k, v in ig_dict.items()] formatted_scores = ", ".join(score_list) # 2. Call Qwen LLM prompt = f"""You are an expert AI auditor interpreting an Explainable AI (XAI) output. A sequence classification model evaluated two policies and predicted their relationship as: {prediction.upper()} Policy A: "{policy_a}" Policy B: "{policy_b}" Token Scores: [{formatted_scores}] Write a highly analytical, 2 to 3 sentence explanation of the model's reasoning. Explicitly quote the specific words that have the highest positive and highest negative scores. Do not hallucinate.""" try: response = llm_client.chat_completion(messages=[{"role": "user", "content": prompt}], max_tokens=1500, temperature=0.1) raw_output = response.choices[0].message.content.strip() # 3. Format the blocks and final output match = re.search(r'(.*?)', raw_output, flags=re.DOTALL) if match: think_content = match.group(1).strip() final_answer = raw_output.replace(match.group(0), '').strip() # --- NEW: TRANSLATE EXPLANATION IF NEEDED --- if lang != "English": think_content = t_text(think_content, lang) final_answer = t_text(final_answer, lang) html_out = f"""
🧠 Click to peek into the AI's thought process
{think_content}
""" return html_out, final_answer, raw_output, ig_json_str if lang != "English": raw_output = t_text(raw_output, lang) return "", raw_output, raw_output, ig_json_str except Exception as e: err_msg = f"⚠️ Explainability Error: {str(e)}" return "", t_text(err_msg, lang) if lang != "English" else err_msg, "", "" def bucket_score(score): """Maps a continuous score [-1.0 to 1.0] to a 7-class drill down.""" if score >= 0.66: return "+3 Indivisible", "coherent" elif score >= 0.33: return "+2 Reinforcing", "coherent" elif score > 0.10: return "+1 Enabling", "coherent" elif score >= -0.10: return "0 Consistent", "neutral" elif score >= -0.33: return "-1 Constraining", "incoherent" elif score >= -0.66: return "-2 Counteracting", "incoherent" else: return "-3 Cancelling", "incoherent" def format_streaming_thoughts(text, is_streaming=True): """Safely formats tags into HTML accordions even before the closing tag arrives.""" if "" not in text: return text formatted = text.replace( "", "
" "🧠 AI is thinking..." "
" ) if "" in formatted: # Close the accordion and change the title once thinking is done formatted = formatted.replace("", "
") formatted = formatted.replace("🧠 AI is thinking...", "🧠 Click to peek into the AI's thought process") formatted = formatted.replace("
{display_lbl}{pct:.1f}%
""" html_parts.append(bar_html) styled_conf = f"
{''.join(html_parts)}
" # JSON string to log purely in the CSV conf_json_str = json.dumps({k: round(v*100, 2) for k, v in prob_dict.items()}) drill_choices = DRILL_DOWN_MAP.get(coarse_label, []) updates.append(( gr.update(value=coarse_label), gr.update(value=styled_conf), gr.update(choices=drill_choices, value=drill_down_label), coarse_label, drill_down_label, conf_json_str )) return updates # ========================================== # 1. AUTHENTICATION (GOOGLE SHEETS VIA SERVICE ACCOUNT) # ========================================== print("Authenticating with Google via Service Account...") gcp_secret = os.environ.get("GCP_CREDENTIALS") if not gcp_secret: raise ValueError("GCP_CREDENTIALS secret not found. Please set it in Hugging Face Space Secrets.") try: creds_dict = json.loads(gcp_secret) gc = gspread.service_account_from_dict(creds_dict) translate_client = translate.Client.from_service_account_info(creds_dict) except json.JSONDecodeError as e: raise ValueError(f"Failed to parse GCP_CREDENTIALS JSON. Error: {e}") spreadsheet = gc.open_by_key('12JM3u10WSpshCcSUEmjhRP5i2bWe9MAK_jrbI56WOCU') def get_worksheet_by_number(spreadsheet, worksheet_number, format=True): worksheet = spreadsheet.get_worksheet(worksheet_number) rows = worksheet.get_all_values() df = pd.DataFrame.from_records(rows[1:], columns=rows[0]) if format: if worksheet_number == 4: df = df.iloc[1:] else: df = df.iloc[2:] df.columns = df.iloc[0].values df = df.iloc[1:] df.columns = [str(col).strip() for col in df.columns] df = df.replace('', pd.NA) if 'Sector' in df.columns: df['Sector'] = df['Sector'].ffill() else: print(f"⚠️ Warning: 'Sector' column missing in worksheet {worksheet_number}. Found columns: {list(df.columns)}") if 'Policy' in df.columns: df['Policy'] = df['Policy'].ffill() return df print("Loading Data from Google Sheets...") land_df = get_worksheet_by_number(spreadsheet, 3, format=True) water_df = get_worksheet_by_number(spreadsheet, 5, format=True) energy_df = get_worksheet_by_number(spreadsheet, 4, format=True) DOMAIN_MAP = {"Land": land_df, "Water": water_df, "Energy": energy_df} DOMAINS = list(DOMAIN_MAP.keys()) # --- EXPERTISE MAPPING --- SECTOR_MAPPING = { "Climate": "Climate", "Water": "Water", "Energy": "Energy", "Land": "Land", "Environment": "Land", "Agriculture": "Land", "Food": "Land", } SECTOR_CHOICES = list(SECTOR_MAPPING.keys()) # ========================================== # 2. CONFIGURATION & TRANSLATION HELPERS # ========================================== HF_TOKEN = os.environ.get("HF_TOKEN") HF_DATASET_REPO = "akaburia/policy-evaluations" HF_CSV_FILE = "policy_coherence_annotations.csv" HF_USERS_FILE = "user_profiles.csv" AVAILABLE_COLUMNS = [ 'Sector', 'Policy', 'General Vision', 'General policy objective', 'Strategic objectives / directions', 'Focus Area / Policy Action Category', 'Policy objectives (of the focus area)', 'Policy Actions and Measures (PAMs)', 'Policy Targets / Indicators' ] DRILL_DOWN_MAP = { "coherent": ["+3 Indivisible", "+2 Reinforcing", "+1 Enabling"], "neutral": ["0 Consistent"], "incoherent": ["-1 Constraining", "-2 Counteracting", "-3 Cancelling"] } MAX_ROWS = 2 LANG_CODES = {"English": "en", "French": "fr", "Portuguese": "pt", "Swahili": "sw"} def t_text(text, target_lang_name): code = LANG_CODES.get(target_lang_name, "en") if code == "en" or not text: return text import html result = translate_client.translate(text, target_language=code) return html.unescape(result["translatedText"]) def t_batch(texts, target_lang_name): code = LANG_CODES.get(target_lang_name, "en") if code == "en" or not texts: return texts import html results = translate_client.translate(texts, target_language=code) return [html.unescape(res["translatedText"]) for res in results] # ========================================== # 3. STANDARD HELPERS # ========================================== def get_unique_items(df, policy_name, col_name): if 'Policy' not in df.columns or col_name not in df.columns: return [] if policy_name not in df['Policy'].values: return [] items = df[df['Policy'] == policy_name][col_name].dropna().unique().tolist() clean_items = [] for i in items: val = str(i).strip() if val and val.lower() not in ['missing', 'nan', 'n/a', 'none', 'null']: clean_items.append(val) return clean_items def get_sector_for_policy(df, policy_name): if 'Policy' not in df.columns or 'Sector' not in df.columns: return "Unknown Sector" if policy_name not in df['Policy'].values: return "Unknown Sector" return str(df[df['Policy'] == policy_name]['Sector'].iloc[0]).strip() def get_policy_list(domain_key): if not domain_key: return [] df = DOMAIN_MAP[domain_key] if 'Policy' not in df.columns: return [] return [p for p in df['Policy'].unique() if pd.notna(p) and str(p).strip()] def load_hf_dataset(): try: path = hf_hub_download(repo_id=HF_DATASET_REPO, filename=HF_CSV_FILE, repo_type="dataset", token=HF_TOKEN) return pd.read_csv(path) except Exception as e: # Added the 3 new Model tracking columns here return pd.DataFrame(columns=[ "Domain_A", "Sector_A", "Policy_A_Name", "Domain_B", "Sector_B", "Policy_B_Name", "Target_Column", "Target_A_Row", "Target_B_Row", "Context_Column", "Context_A_Chunk", "Context_B_Chunk", "Model_Coarse_Label", "Model_Drill_Down_Label", "Model_Confidences", "AI_Justification", "IG_JSON", "Coherence_Label", "Drill_Down_Label", "Justification", "AnnotatorUsername", "Timestamp", "SessionID", "Consent_Link_Email", "Consent_Follow_Up" ]) def load_user_profiles(): try: path = hf_hub_download(repo_id=HF_DATASET_REPO, filename=HF_USERS_FILE, repo_type="dataset", token=HF_TOKEN) return pd.read_csv(path) except Exception: return pd.DataFrame(columns=["Email", "UserID"]) def get_or_create_user(email): email = email.strip().lower() if not email: return None, "Email cannot be empty." users_df = load_user_profiles() if email in users_df['Email'].values: user_id = users_df.loc[users_df['Email'] == email, 'UserID'].iloc[0] return user_id, f"Welcome back. Logged in as {user_id}." else: new_num = len(users_df) + 1 new_user_id = f"user{new_num}" new_row = {"Email": email, "UserID": new_user_id} users_df = pd.concat([users_df, pd.DataFrame([new_row])], ignore_index=True) try: csv_buffer = io.StringIO() users_df.to_csv(csv_buffer, index=False) api = HfApi() api.upload_file( path_or_fileobj=io.BytesIO(csv_buffer.getvalue().encode('utf-8')), path_in_repo=HF_USERS_FILE, repo_id=HF_DATASET_REPO, token=HF_TOKEN, repo_type="dataset" ) return new_user_id, f"New account created. Logged in as {new_user_id}." except Exception as e: return None, f"Error saving user profile: {e}" # ========================================== # 4. GRADIO UI DESIGN # ========================================== custom_css = """ .explain-btn { background-color: #8b5cf6 !important; color: white !important; border: none !important; } .explain-btn:hover { background-color: #7c3aed !important; } .scrollable-target textarea { min-height: 80px !important; overflow-y: auto !important; } .scrollable-rows-container { padding: 5px !important; background-color: #f9fafb !important; } /* Clean card layout for individual rows */ .row-card { padding: 15px !important; background: #ffffff !important; border-radius: 8px !important; box-shadow: 0 1px 3px rgba(0,0,0,0.1) !important; margin-bottom: 20px !important; border: 1px solid #e5e7eb !important; } """ with gr.Blocks(theme=gr.themes.Soft(), css=custom_css) as demo: # --- TOP BANNER --- gr.HTML("""
""") # --- PERSISTENT HEADER --- with gr.Row(): with gr.Column(scale=4): main_title = gr.HTML("""

Policy Coherence Tool

""") with gr.Column(scale=1): gr.HTML("""
""") lang_selector = gr.Dropdown(choices=list(LANG_CODES.keys()), value="English", label="Language / Langue", scale=1) # --- 1. LANDING PAGE --- with gr.Column(visible=True) as landing_page: gr.HTML("""
""") main_desc = gr.Markdown( "### Mapping Policy Synergies Across Sectors\n\n" "Welcome to the EPIC Africa Policy Coherence Tool. This platform is designed to help researchers and policymakers " "systematically evaluate how specific objectives within different policy documents interact with one another.\n\n" "**How it works:**\n" "You will be presented with a target objective from one policy and asked to score its interaction against objectives from a different policy. " "By identifying whether these targets reinforce, enable, or constrain one another, you help build a comprehensive understanding of cross-sectoral coherence.\n\n" "**Your contribution matters:**\n" ) get_started_btn = gr.Button("Get Started", variant="primary", size="lg") hf_df_state = gr.State() user_tag_state = gr.State() session_id_state = gr.State(lambda: str(uuid.uuid4().hex[:12])) consent_link_state = gr.State("No") consent_follow_state = gr.State("No") target_a_list_state = gr.State([]) pending_tasks_state = gr.State({}) current_index_state = gr.State(0) current_b_eng_list_state = gr.State([]) ctx_a_eng_state = gr.State("") ctx_b_eng_state = gr.State("") # --- LOGIN PANEL --- with gr.Group(visible=False) as login_box: login_title = gr.Markdown("### User Authentication & Informed Consent") login_disclaimer = gr.Markdown( "**Before you continue**\n\n" "This survey is anonymous by default. When you sign in with your email, we will ask two brief questions about how your data is handled.\n\n" "**1 · Linking your responses to your email**\n" "You may choose to have your responses stored in association with your email address. This is entirely optional. Your data will be held securely and will not be shared with any third party.\n\n" "**2 · Follow-up contact**\n" "If you agreed to linking above, we may also ask whether you are willing to be contacted for a brief follow-up conversation — only in cases where your responses raise questions that would benefit from further discussion.\n\n" "*You can say no to either question without affecting your participation.*" ) with gr.Row(): consent_link_radio = gr.Radio(choices=["Yes", "No"], value="No", label="1. Link responses to my email?") consent_follow_radio = gr.Radio(choices=["Yes", "No"], value="No", label="2. Willing to be contacted for follow-up?", interactive=False) with gr.Row(): email_box = gr.Textbox(label="Email Address", placeholder="name@example.com") login_btn = gr.Button("Login & Accept", variant="primary") login_status = gr.Markdown(value="Waiting for authentication...") # Logic to disable the second question if the first is 'No' def toggle_followup(link_choice): if link_choice == "Yes": return gr.update(interactive=True) else: return gr.update(value="No", interactive=False) consent_link_radio.change(fn=toggle_followup, inputs=consent_link_radio, outputs=consent_follow_radio) # --- EXPERTISE / SECTOR SELECTION --- with gr.Group(visible=False) as sector_box: sector_title = gr.Markdown("### What is your expertise?") sector_cb = gr.CheckboxGroup( choices=SECTOR_CHOICES, label="Please select the sector(s) that best match your expertise and work experience. Multiple selections are allowed." ) proceed_btn = gr.Button("Proceed to Workspace", variant="primary", size="lg") # --- MAIN APPLICATION --- with gr.Group(visible=False) as app_box: app_definitions = gr.Markdown( "**Definitions:**\n" "- **Nexus Domain:** The broad sector being analyzed (e.g., Land, Water, Energy).\n" "- **Policy:** The specific document under review.\n" "- **Target:** The exact objective or statement you are currently evaluating.\n" "- **Context:** The broader set of measures belonging to the policy, provided as background reference.\n\n" "**General Class Definitions:**\n" "- **Supportive:** the policy objectives explicitly reinforce each other\n" "- **Non-interacting:** policy objectives are independent of each other and non-related\n" "- **Contradictory:** the policy objectives imply no explicit alignment or are contradicting of each other" ) # NEW: Accordion containing the 7 classes table with gr.Accordion("Interaction Class Definitions (Click to Expand)", open=False) as interaction_acc: interaction_md = gr.Markdown( "| Interaction Label | Meaning | Implication |\n" "| :--- | :--- | :--- |\n" "| **+3 (Indivisible)** | Progress on one target automatically delivers progress on another | There is high level of compatibility between the two targets. |\n" "| **+2 (Reinforcing)** | Progress on one target makes it easier to make progress on another | There is relatively higher level of compatibility between the targets being compared. |\n" "| **+1 (Enabling)** | Progress on one target creates conditions that enable progress on another | There is a small level of compatibility between the two targets compared. |\n" "| **0 (Consistent)** | There is no significant link between two targets' progress | There is no significant compatibility between the two targets being evaluated. |\n" "| **-1 (Constraining)** | Progress on one target constrains the options for how to deliver on another | The targets are relatively competitive resulting in counterproductive effects. |\n" "| **-2 (Counteracting)** | Progress on one target makes it more difficult to make progress on another | The targets are counterproductive and do not support each other. |\n" "| **-3 (Cancelling)** | Progress on one target automatically leads to a negative impact on another | The targets are highly opposite and are highly counterproductive. Cannot deliver synergistic effects. |" ) with gr.Accordion("Data Selection", open=True) as data_acc: # --- NEW LOCATION FOR BACK BUTTON --- with gr.Row(): back_to_sectors_btn = gr.Button("⬅️ Back to Sector Selection", variant="secondary", size="sm") gr.Markdown("") # Empty markdown to push button to the left gr.Markdown("") with gr.Row(): with gr.Column(scale=1): src_a_title = gr.Markdown("### Source A") domain_a_dd = gr.Dropdown(choices=DOMAINS, value=None, label="Domain A") policy_a_dd = gr.Dropdown(choices=[], value=None, label="Policy A") with gr.Column(scale=1): src_b_title = gr.Markdown("### Source B") domain_b_dd = gr.Dropdown(choices=DOMAINS, value=None, label="Domain B") policy_b_dd = gr.Dropdown(choices=[], value=None, label="Policy B") with gr.Row(): target_col_dd = gr.Dropdown(choices=AVAILABLE_COLUMNS, value='Strategic objectives / directions', label="Target Column") context_col_dd = gr.Dropdown(choices=AVAILABLE_COLUMNS, value='Policy Actions and Measures (PAMs)', label="Context Column") # Just the Load Button at the bottom now load_btn = gr.Button("Fetch Data", variant="primary") gr.Markdown("---") progress_text = gr.Markdown("**Progress:** Waiting for data selection...") with gr.Group(visible=False) as workspace_box: # --- THE FIXED HEADER --- # This stays naturally at the top of the workspace with gr.Row(): with gr.Column(scale=1, variant="panel"): meta_a = gr.Markdown("### Source A Information") display_context_a = gr.Textbox(label="Context A", interactive=False, lines=4) # display_target_a = gr.Textbox(label="Target A (Active)", interactive=False, lines=4) with gr.Column(scale=1, variant="panel"): meta_b = gr.Markdown("### Source B Information") display_context_b = gr.Textbox(label="Context B", interactive=False, lines=4) # --- AI CHATBOT QUERY OPTION --- with gr.Accordion("💬 Ask AI about the Context & Policies", open=False): chatbot = gr.Chatbot(height=300) with gr.Row(): chat_input = gr.Textbox(placeholder="Ask a question about the policies or targets...", scale=4, show_label=False) chat_submit = gr.Button("Send", scale=1) # --- THE SCROLLABLE ROWS --- with gr.Group(elem_classes="scrollable-rows-container"): bulk_title = gr.Markdown("### Bulk Coherence Evaluation") bulk_desc = gr.Markdown( "Evaluate how the **Target A** above interacts with the **Target B** statements below.\n" "**Rules:** If you evaluate a row, you **MUST** select the Class, the Extended Class Interaction, and write a Justification. " "You may leave a row entirely blank to skip it." ) # --- DYNAMIC BULK ROWS --- eval_rows = [] for i in range(MAX_ROWS): with gr.Group(visible=False, elem_classes="row-card") as row_container: m_coarse_st = gr.State("") m_drill_st = gr.State("") m_conf_st = gr.State("") m_ai_just_st = gr.State("") m_ig_json_st = gr.State("") # FIX: Show Target A and Target B side-by-side in every row block with gr.Row(equal_height=True): with gr.Column(scale=1): a_text_display = gr.Textbox(label="Target A (Active)", interactive=False, lines=3, elem_classes="scrollable-target") with gr.Column(scale=1): b_text = gr.Textbox(label="Target B", interactive=False, lines=3, elem_classes="scrollable-target") with gr.Row(): with gr.Column(scale=1, min_width=200): rel_radio = gr.Radio(choices=[("Supportive", "coherent"), ("Non-interacting", "neutral"), ("Contradictory", "incoherent")], label="1. Class") conf_md = gr.Markdown("") with gr.Column(scale=1, min_width=200): # ADDED allow_custom_value=True to prevent strict validation crashes on swap/clear inter_dd = gr.Dropdown(choices=[], label="2. Extended Class Interaction", interactive=True, allow_custom_value=True) explain_btn = gr.Button("✨ AI Explainability", size="sm", elem_classes="explain-btn") explain_html = gr.HTML("") with gr.Column(scale=2, min_width=250): just_box = gr.Textbox(label="3. Justification", placeholder="Compulsory reasoning...", lines=3) clear_row_btn = gr.Button("🗑️ Clear", size="sm", variant="stop") explain_btn.click( fn=generate_row_explanation, inputs=[target_a_list_state, current_index_state, b_text, lang_selector], # Passed lang_selector here outputs=[explain_html, just_box, m_ai_just_st, m_ig_json_st] ) clear_row_btn.click( fn=lambda: (gr.update(value=None), gr.update(choices=[], value=None), gr.update(value="")), inputs=None, outputs=[rel_radio, inter_dd, just_box] ) # Added a_text_display to the tuple eval_rows.append((row_container, a_text_display, b_text, rel_radio, conf_md, inter_dd, just_box, m_coarse_st, m_drill_st, m_conf_st, m_ai_just_st, m_ig_json_st)) # --- NAVIGATION BUTTONS --- with gr.Row(): skip_btn = gr.Button("Skip Target A", size="lg") save_btn = gr.Button("Save Filled Annotations", variant="primary", size="lg") status_box = gr.Textbox(label="System Log", interactive=False) workspace_info = gr.Markdown( "
" "Visualisations Report
" "As you keep on scoring, head on to check out the visualisations report! This is an exercise on scoring policy coherences and comparison with the AI model.
" "https://datastudio.google.com/u/0/reporting/cc1d6cab-fe9d-4d77-91d9-ea6ccfb5a39c/page/U7RuF" "
" ) # --- NAVIGATION BUTTONS --- # Kept outside the scrollable box so they are always visible at the very bottom # with gr.Row(): # skip_btn = gr.Button("Skip Target A", size="lg") # save_btn = gr.Button("Save Filled Annotations", variant="primary", size="lg") # status_box = gr.Textbox(label="System Log", interactive=False) # ========================================== # 5. EVENT CONTROLLERS # ========================================== # --- PERSISTENT FOOTER --- footer_disclaimer = gr.Markdown( "---\n" "
" "Disclaimer: This tool is developed and maintained by EPIC Africa. " "The European Union (EU) is not liable for the content, use, or outputs generated by this tool." "
" ) gr.HTML(""" """) def translate_static_ui(lang): titles = [ """

Policy Coherence Tool

""", "### Mapping Policy Synergies Across Sectors\n\nWelcome to the EPIC Africa Policy Coherence Tool. This platform is designed to help researchers and policymakers systematically evaluate how specific objectives within different policy documents interact with one another.\n\n**How it works:**\nYou will be presented with a target objective from one policy and asked to score its interaction against objectives from a different policy. By identifying whether these targets reinforce, enable, or constrain one another, you help build a comprehensive understanding of cross-sectoral coherence.\n\n", "Get Started", "### User Authentication & Informed Consent", "**Before you continue**\n\nThis survey is anonymous by default. When you sign in with your email, we will ask two brief questions about how your data is handled.\n\n**1 · Linking your responses to your email**\nYou may choose to have your responses stored in association with your email address. This is entirely optional. Your data will be held securely and will not be shared with any third party.\n\n**2 · Follow-up contact**\nIf you agreed to linking above, we may also ask whether you are willing to be contacted for a brief follow-up conversation — only in cases where your responses raise questions that would benefit from further discussion.\n\n*You can say no to either question without affecting your participation.*", "Login & Accept", "### What is your expertise?", "Please select the sector(s) that best match your expertise and work experience. Multiple selections are allowed.", "Proceed to Workspace", "**Definitions:**\n- **Nexus Domain:** The broad sector being analyzed (e.g., Land, Water, Energy).\n- **Policy:** The specific document under review.\n- **Target:** The exact objective or statement you are currently evaluating.\n- **Context:** The broader set of measures belonging to the policy, provided as background reference.\n\n**General Class Definitions:**\n- **Supportive:** the policy objectives explicitly reinforce each other\n- **Non-interacting:** policy objectives are independent of each other and non-related\n- **Contradictory:** the policy objectives imply no explicit alignment or are contradicting of each other", "Interaction Class Definitions (Click to Expand)", "| Interaction Label | Meaning | Implication |\n| :--- | :--- | :--- |\n| **+3 (Indivisible)** | Progress on one target automatically delivers progress on another | There is high level of compatibility between the two targets. |\n| **+2 (Reinforcing)** | Progress on one target makes it easier to make progress on another | There is relatively higher level of compatibility between the targets being compared. |\n| **+1 (Enabling)** | Progress on one target creates conditions that enable progress on another | There is a small level of compatibility between the two targets compared. |\n| **0 (Consistent)** | There is no significant link between two targets' progress | There is no significant compatibility between the two targets being evaluated. |\n| **-1 (Constraining)** | Progress on one target constrains the options for how to deliver on another | The targets are relatively competitive resulting in counterproductive effects. |\n| **-2 (Counteracting)** | Progress on one target makes it more difficult to make progress on another | The targets are counterproductive and do not support each other. |\n| **-3 (Cancelling)** | Progress on one target automatically leads to a negative impact on another | The targets are highly opposite and are highly counterproductive. Cannot deliver synergistic effects. |", "Data Selection", "⬅️ Back to Sector Selection", "### Source A", "### Source B", "Fetch Data", "💬 Ask AI about the Context & Policies", "### Bulk Coherence Evaluation", "Evaluate how the **Target A** above interacts with the **Target B** statements below.\n**Rules:** If you evaluate a row, you **MUST** select the Class, the Extended Class Interaction, and write a Justification. You may leave a row entirely blank to skip it.", "Skip Target A", "Save Filled Annotations", "
📊 Visualisations Report
Watch the dataset grow as we map policy coherences across domains! Check out the live visualisations report to track our collective progress, uncover cross-sector synergies, and see exactly how our policy experts stack up against the AI baseline.
https://datastudio.google.com/u/0/reporting/cc1d6cab-fe9d-4d77-91d9-ea6ccfb5a39c/page/U7RuF
", "---\n
Disclaimer: This tool is developed and maintained by EPIC Africa. The European Union (EU) is not liable for the content, use, or outputs generated by this tool.
" ] translated = t_batch(titles, lang) return translated def handle_language_change(lang, ctx_a_eng, ctx_b_eng, a_list, tasks_dict, idx, user_tag, pol_a, pol_b, hf_df): static_updates = translate_static_ui(lang) ctx_a_trans = t_text(ctx_a_eng, lang) ctx_b_trans = t_text(ctx_b_eng, lang) rendered = render_target_a(a_list, tasks_dict, idx, lang, user_tag, pol_a, pol_b, hf_df) prog_txt = rendered[0] row_updates = rendered[1:] return [ gr.update(value=static_updates[0]), # main_title HTML gr.update(value=static_updates[1]), # main_desc gr.update(value=static_updates[2]), # get_started_btn gr.update(value=static_updates[3]), # login_title gr.update(value=static_updates[4]), # login disclaimer gr.update(value=static_updates[5]), # login_btn gr.update(value=static_updates[6]), # sector_title gr.update(label=static_updates[7]), # sector_cb label gr.update(value=static_updates[8]), # proceed_btn gr.update(value=static_updates[9]), # app_definitions gr.update(label=static_updates[10]), # interaction_acc gr.update(value=static_updates[11]), # interaction_md gr.update(label=static_updates[12]), # data_acc gr.update(value=static_updates[13]), # back_to_sectors_btn gr.update(value=static_updates[14]), # src_a_title gr.update(value=static_updates[15]), # src_b_title gr.update(value=static_updates[16]), # load_btn # Missing chat accordion translation mapping index, let's omit the chat accordion title for simplicity since we can't easily grab it. gr.update(value=static_updates[18]), # bulk_title gr.update(value=static_updates[19]), # bulk_desc gr.update(value=static_updates[20]), # skip_btn gr.update(value=static_updates[21]), # save_btn gr.update(value=static_updates[22]), # workspace_info gr.update(value=static_updates[23]), # footer_disclaimer gr.update(value=ctx_a_trans), # ctx_a_trans gr.update(value=ctx_b_trans), # ctx_b_trans prog_txt ] + row_updates def update_drill(label, current_val): # Gracefully handle the clear button event to avoid validation errors if not label: return gr.update(choices=[], value=None) choices = DRILL_DOWN_MAP.get(label, []) if current_val in choices: return gr.update(choices=choices, value=current_val) new_val = choices[0] if choices else None return gr.update(choices=choices, value=new_val) # for i in range(MAX_ROWS): # # Change this line to have 11 items (add two more underscores at the end) # _, _, rel_radio, _, inter_dd, _, _, _, _, _, _ = eval_rows[i] # # Notice we now pass BOTH the radio and the dropdown as inputs # rel_radio.change(fn=update_drill, inputs=[rel_radio, inter_dd], outputs=inter_dd) for i in range(MAX_ROWS): # FIX: Added an extra '_' at the beginning to properly unpack 12 items instead of 11 _, _, b_text, rel_radio, _, inter_dd, just_box, _, _, _, _, _ = eval_rows[i] # Restore the missing drill-down update event rel_radio.change(fn=update_drill, inputs=[rel_radio, inter_dd], outputs=inter_dd) # Gather the exact state needed to cache this row AND the workspace config inputs_to_cache = [ user_tag_state, session_id_state, domain_a_dd, policy_a_dd, domain_b_dd, policy_b_dd, target_col_dd, context_col_dd, target_a_list_state, current_index_state, b_text, rel_radio, inter_dd, just_box ] # Trigger cache save silently in the background on any change rel_radio.change(fn=update_cache_row, inputs=inputs_to_cache) inter_dd.change(fn=update_cache_row, inputs=inputs_to_cache) just_box.change(fn=update_cache_row, inputs=inputs_to_cache) # ADD link_val and follow_val to inputs def authenticate(email, link_val, follow_val): email = email.strip().lower() # 1. Validate Email Format email_pattern = r"^[^@\s]+@[^@\s]+\.[^@\s]+$" if not re.match(email_pattern, email): write_log("LOGIN_FAILED", f"Invalid email format attempted: '{email}'") return (gr.update(value=f"Please enter a valid email address."), gr.update(visible=True), gr.update(visible=False), gr.update(visible=False), None, None, gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), link_val, follow_val) user_tag, msg = get_or_create_user(email) if not user_tag: write_log("LOGIN_FAILED", f"System error creating user for '{email}'") return (gr.update(value=f"{msg}"), gr.update(visible=True), gr.update(visible=False), gr.update(visible=False), None, None, gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), link_val, follow_val) write_log("LOGIN_SUCCESS", f"User {user_tag} ({email}) logged in. Consents - Link: {link_val}, Follow: {follow_val}") hf_df = load_hf_dataset() drafts = load_drafts() user_data = drafts.get(user_tag, {}) ws = user_data.get("workspace", {}) # 2. Check for pending session and redirect straight to workspace if ws.get("pol_a") and ws.get("pol_b"): write_log("SESSION_RESTORED", f"User {user_tag} skipped sectors and restored previous session {ws.get('session_id')}.") msg += f" Restored your pending session. Click 'Fetch Data' to resume your draft." return ( gr.update(value=f"{msg} Loaded {len(hf_df)} annotations."), gr.update(visible=False), # Hide login_box gr.update(visible=False), # Hide sector_box (Bypass directly to app) gr.update(visible=True), # Show app_box user_tag, hf_df, gr.update(value=ws["dom_a"]), gr.update(choices=get_policy_list(ws["dom_a"]), value=ws["pol_a"]), gr.update(value=ws["dom_b"]), gr.update(choices=get_policy_list(ws["dom_b"]), value=ws["pol_b"]), gr.update(value=ws["tar_col"]), gr.update(value=ws["ctx_col"]), link_val, follow_val ) else: write_log("NEW_SESSION", f"User {user_tag} starting fresh. Routing to Sector Selection.") return ( gr.update(value=f"{msg} Loaded {len(hf_df)} annotations."), gr.update(visible=False), # Hide login_box gr.update(visible=True), # Show sector_box gr.update(visible=False), # Hide app_box user_tag, hf_df, gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), link_val, follow_val ) def route_to_workspace(selected_sectors): if not selected_sectors: raise gr.Error("Please select at least one sector.") allowed_domains = set() for s in selected_sectors: mapped_domain = SECTOR_MAPPING.get(s) if mapped_domain in DOMAINS: allowed_domains.add(mapped_domain) allowed_list = list(allowed_domains) # Pre-select the first domain, but dynamically load its corresponding policies default_domain = allowed_list[0] if allowed_list else None available_policies = get_policy_list(default_domain) if default_domain else [] write_log("SECTOR_SELECTED", f"Mapped sectors {selected_sectors} to domains {allowed_list}") return ( gr.update(visible=False), # Hide sector box gr.update(visible=True), # Show main app gr.update(choices=allowed_list, value=default_domain), # Restrict Domain A gr.update(choices=available_policies, value=None), # <--- POPULATE Policy A choices, leave value empty gr.update(choices=DOMAINS, value=DOMAINS[0] if DOMAINS else None), # Open Domain B gr.update(choices=get_policy_list(DOMAINS[0]) if DOMAINS else [], value=None) # <--- POPULATE Policy B choices, leave value empty ) def update_a_choices(dom_a, pol_b, curr_a): choices = [p for p in get_policy_list(dom_a) if p != pol_b] val = curr_a if curr_a in choices else None return gr.update(choices=choices, value=val) def update_b_choices(dom_b, pol_a, curr_b): choices = [p for p in get_policy_list(dom_b) if p != pol_a] val = curr_b if curr_b in choices else None return gr.update(choices=choices, value=val) get_started_btn.click( fn=lambda: (gr.update(visible=False), gr.update(visible=True)), inputs=None, outputs=[landing_page, login_box] ) login_btn.click( fn=authenticate, inputs=[email_box, consent_link_radio, consent_follow_radio], outputs=[ login_status, login_box, sector_box, app_box, user_tag_state, hf_df_state, # <-- Added sector_box here domain_a_dd, policy_a_dd, domain_b_dd, policy_b_dd, target_col_dd, context_col_dd, consent_link_state, consent_follow_state ] ) # --- WIRE THE NEW PROCEED BUTTON --- proceed_btn.click( fn=route_to_workspace, inputs=[sector_cb], outputs=[sector_box, app_box, domain_a_dd, policy_a_dd, domain_b_dd, policy_b_dd] # <-- Added policy dropdowns ) back_to_sectors_btn.click( fn=lambda: (gr.update(visible=True), gr.update(visible=False)), inputs=None, outputs=[sector_box, app_box] ) domain_a_dd.change(fn=update_a_choices, inputs=[domain_a_dd, policy_b_dd, policy_a_dd], outputs=policy_a_dd) policy_b_dd.change(fn=update_a_choices, inputs=[domain_a_dd, policy_b_dd, policy_a_dd], outputs=policy_a_dd) domain_b_dd.change(fn=update_b_choices, inputs=[domain_b_dd, policy_a_dd, policy_b_dd], outputs=policy_b_dd) policy_a_dd.change(fn=update_b_choices, inputs=[domain_b_dd, policy_a_dd, policy_b_dd], outputs=policy_b_dd) def render_target_a(a_list, tasks_dict, idx, lang, user_tag, pol_a, pol_b, hf_df, progress=gr.Progress()): updates = [] empty_row = [gr.update(visible=False), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), "", "", "", "", ""] if not a_list: prog_txt = t_text("**Progress:** No unannotated items found.", lang) updates.append(prog_txt) for i in range(MAX_ROWS): updates.extend(empty_row) return updates + [[]] if idx >= len(a_list): prog_txt = t_text("**Progress:** Completed all items.", lang) updates.append(prog_txt) for i in range(MAX_ROWS): updates.extend(empty_row) return updates + [[]] curr_a_eng = a_list[idx] bs_to_eval_eng = tasks_dict[curr_a_eng] curr_a_display = t_text(curr_a_eng, lang) bs_display = t_batch(bs_to_eval_eng, lang) prog_txt = t_text(f"**Progress:** Annotating Target A group {idx + 1} of {len(a_list)}", lang) updates.append(prog_txt) drafts = load_drafts() cache_key = f"{pol_a}|{pol_b}|{curr_a_eng}" user_draft = drafts.get(user_tag, {}).get("rows", {}).get(cache_key, {}) user_saved_df = pd.DataFrame() if not hf_df.empty: temp_df = hf_df[ (hf_df["AnnotatorUsername"] == user_tag) & (hf_df["Policy_A_Name"] == pol_a) & (hf_df["Policy_B_Name"] == pol_b) & (hf_df["Target_A_Row"] == curr_a_eng) ].copy() if not temp_df.empty: temp_df['Timestamp'] = pd.to_datetime(temp_df['Timestamp']) temp_df = temp_df.sort_values(by='Timestamp') user_saved_df = temp_df.drop_duplicates(subset=["Target_B_Row"], keep="last") # --- PROGRESS BAR UPDATE --- if progress is not None: progress(0.4, desc="Running background AI predictions...") preds = get_model_predictions(curr_a_eng, bs_to_eval_eng) # --- PROGRESS BAR UPDATE --- if progress is not None: progress(0.8, desc="Rendering UI blocks...") for i in range(MAX_ROWS): if i < len(bs_display): p_radio, p_conf_md, p_inter_dd, p_m_coarse, p_m_drill, p_m_conf = preds[i] b_val_eng = bs_to_eval_eng[i] cached_row = user_draft.get(b_val_eng) saved_row = user_saved_df[user_saved_df["Target_B_Row"] == b_val_eng] if not user_saved_df.empty else pd.DataFrame() if cached_row: set_radio = gr.update(value=cached_row.get("rel")) if cached_row.get("rel") else p_radio set_inter = gr.update(value=cached_row.get("inter")) if cached_row.get("inter") else p_inter_dd set_just = gr.update(value=cached_row.get("just", "")) elif not saved_row.empty: set_radio = gr.update(value=saved_row.iloc[-1]["Coherence_Label"]) set_inter = gr.update(value=saved_row.iloc[-1]["Drill_Down_Label"]) set_just = gr.update(value=saved_row.iloc[-1]["Justification"]) else: set_radio = p_radio set_inter = p_inter_dd set_just = gr.update(value="") updates.extend([ gr.update(visible=True), gr.update(value=curr_a_display), gr.update(value=bs_display[i]), set_radio, p_conf_md, set_inter, set_just, p_m_coarse, p_m_drill, p_m_conf, "", "" ]) else: updates.extend(empty_row) # --- PROGRESS BAR UPDATE --- if progress is not None: progress(1.0, desc="Done!") return updates + [bs_to_eval_eng] # def render_target_a(a_list, tasks_dict, idx, lang): # updates = [] # # 9 components per row to reset # # empty_row = [gr.update(visible=False), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), "", "", ""] # empty_row = [gr.update(visible=False), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), "", "", "", "", ""] # if not a_list: # prog_txt = t_text("**Progress:** No unannotated items found.", lang) # updates.extend([prog_txt, "N/A"]) # for i in range(MAX_ROWS): updates.extend(empty_row) # return updates + [[]] # if idx >= len(a_list): # prog_txt = t_text("**Progress:** Completed all items.", lang) # updates.extend([prog_txt, t_text("End of list.", lang)]) # for i in range(MAX_ROWS): updates.extend(empty_row) # return updates + [[]] # curr_a_eng = a_list[idx] # bs_to_eval_eng = tasks_dict[curr_a_eng] # curr_a_display = t_text(curr_a_eng, lang) # bs_display = t_batch(bs_to_eval_eng, lang) # prog_txt = t_text(f"**Progress:** Annotating Target A group {idx + 1} of {len(a_list)}", lang) # updates.extend([prog_txt, curr_a_display]) # for i in range(MAX_ROWS): # if i < len(bs_display): # updates.extend([ # gr.update(visible=True), # gr.update(value=bs_display[i]), # gr.update(value=None), # gr.update(value=""), # conf_md # gr.update(choices=[], value=None), # gr.update(value=""), # just_box # "", "", "", "", "" # Reset the 5 hidden model states # ]) # else: # updates.extend(empty_row) # return updates + [bs_to_eval_eng] def load_workspace(dom_a, pol_a, dom_b, pol_b, tar_col, ctx_col, hf_df, user_tag, lang, progress=gr.Progress()): progress(0.1, desc="Validating selections...") if not pol_a or not pol_b: err = t_text("Error: Select both policies.", lang) return [gr.update(value=err)] + [gr.skip()] * (14 + MAX_ROWS*12) if tar_col == ctx_col: err = t_text("Error: Target and Context cannot be the same.", lang) return [gr.update(value=err)] + [gr.skip()] * (14 + MAX_ROWS*12) progress(0.2, desc="Extracting policy structures...") df_a = DOMAIN_MAP[dom_a] df_b = DOMAIN_MAP[dom_b] sec_a = get_sector_for_policy(df_a, pol_a) sec_b = get_sector_for_policy(df_b, pol_b) meta_a_str = f"**Sector:** {sec_a} | **Policy:** {pol_a}" meta_b_str = f"**Sector:** {sec_b} | **Policy:** {pol_b}" targets_a = get_unique_items(df_a, pol_a, tar_col) targets_b = get_unique_items(df_b, pol_b, tar_col) user_df = hf_df[hf_df["AnnotatorUsername"] == user_tag] mask = (user_df["Policy_A_Name"] == pol_a) & (user_df["Policy_B_Name"] == pol_b) annotated_pairs = set(zip(user_df.loc[mask, "Target_A_Row"], user_df.loc[mask, "Target_B_Row"])) pending_tasks = {} total_missing_pairs = 0 for a in targets_a: missing_bs = [b for b in targets_b if (a, b) not in annotated_pairs] if missing_bs: pending_tasks[a] = missing_bs[:MAX_ROWS] total_missing_pairs += len(pending_tasks[a]) target_a_list = list(pending_tasks.keys()) contexts_a = get_unique_items(df_a, pol_a, ctx_col) contexts_b = get_unique_items(df_b, pol_b, ctx_col) ctx_a_chunk_eng = "\n\n".join([f"• {c}" for c in contexts_a]) if contexts_a else "No context data available." ctx_b_chunk_eng = "\n\n".join([f"• {c}" for c in contexts_b]) if contexts_b else "No context data available." ctx_a_display = t_text(ctx_a_chunk_eng, lang) ctx_b_display = t_text(ctx_b_chunk_eng, lang) rendered_updates = render_target_a(target_a_list, pending_tasks, 0, lang, user_tag, pol_a, pol_b, hf_df, progress) prog = rendered_updates[0] row_updates = rendered_updates[1:-1] b_eng_list = rendered_updates[-1] status_msg = t_text(f"Data loaded. {total_missing_pairs} unannotated pairs remain across {len(target_a_list)} Target A groups.", lang) write_log("WORKSPACE_LOADED", f"User {user_tag} fetched {pol_a} vs {pol_b}.") return [ target_a_list, pending_tasks, 0, ctx_a_chunk_eng, ctx_b_chunk_eng, b_eng_list, prog, meta_a_str, ctx_a_display, meta_b_str, ctx_b_display, status_msg, gr.update(visible=len(target_a_list) > 0) ] + row_updates def save_action(idx, a_list, tasks_dict, ctx_a_chunk_eng, ctx_b_chunk_eng, b_eng_list, dom_a, pol_a, dom_b, pol_b, tar_col, ctx_col, user_tag, session_id, c_link, c_follow, hf_df, lang, *row_data): if idx >= len(a_list): return gr.update(value=t_text("End of list reached.", lang)), idx, hf_df current_a_eng = a_list[idx] new_rows = [] # Generate exact local timestamp try: tz = ZoneInfo("Africa/Nairobi") except: import pytz tz = pytz.timezone("Africa/Nairobi") current_time = datetime.now(tz).strftime('%Y-%m-%d %H:%M:%S') for i in range(MAX_ROWS): if i >= len(b_eng_list): break b_val_eng = b_eng_list[i] rel = row_data[i*10 + 1] inter = row_data[i*10 + 3] just = row_data[i*10 + 4] model_coarse = row_data[i*10 + 5] model_drill = row_data[i*10 + 6] model_conf = row_data[i*10 + 7] ai_just = row_data[i*10 + 8] ig_json = row_data[i*10 + 9] has_rel = bool(rel) has_inter = bool(inter) has_just = bool(just and just.strip()) if has_rel or has_inter or has_just: if not (has_rel and has_inter and has_just): raise gr.Error(f"Row {i+1} is incomplete! Please fill Class, Extended Class, and Justification, or clear the row to skip.") new_rows.append({ "Domain_A": dom_a, "Sector_A": get_sector_for_policy(DOMAIN_MAP[dom_a], pol_a), "Policy_A_Name": pol_a, "Domain_B": dom_b, "Sector_B": get_sector_for_policy(DOMAIN_MAP[dom_b], pol_b), "Policy_B_Name": pol_b, "Target_Column": tar_col, "Target_A_Row": current_a_eng, "Target_B_Row": b_val_eng, "Context_Column": ctx_col, "Context_A_Chunk": ctx_a_chunk_eng, "Context_B_Chunk": ctx_b_chunk_eng, "Model_Coarse_Label": model_coarse, "Model_Drill_Down_Label": model_drill, "Model_Confidences": model_conf, "AI_Justification": ai_just, "IG_JSON": ig_json, "Coherence_Label": rel, "Drill_Down_Label": inter, "Justification": just.strip(), "AnnotatorUsername": user_tag, "Timestamp": current_time, "SessionID": session_id, "Consent_Link_Email": c_link, "Consent_Follow_Up": c_follow }) if new_rows: new_df = pd.DataFrame(new_rows) hf_df = pd.concat([hf_df, new_df], ignore_index=True) try: csv_buffer = io.StringIO() hf_df.to_csv(csv_buffer, index=False) csv_bytes = csv_buffer.getvalue().encode('utf-8') write_log("DATA_SAVED", f"User {user_tag} successfully saved {len(new_rows)} completed rows to Hugging Face.") api = HfApi() api.upload_file( path_or_fileobj=io.BytesIO(csv_bytes), path_in_repo=HF_CSV_FILE, repo_id=HF_DATASET_REPO, token=HF_TOKEN, repo_type="dataset" ) log_msg = t_text(f"Successfully saved {len(new_rows)} annotations.", lang) # CLEAR CACHE ON SUCCESSFUL SAVE drafts = load_drafts() cache_key = f"{pol_a}|{pol_b}|{current_a_eng}" # Check inside the "rows" sub-dictionary if user_tag in drafts and "rows" in drafts[user_tag] and cache_key in drafts[user_tag]["rows"]: del drafts[user_tag]["rows"][cache_key] with open(DRAFT_FILE, 'w') as f: json.dump(drafts, f) except Exception as e: log_msg = f"Error saving data: {e}" else: log_msg = t_text("No annotations filled. Skipped to next group.", lang) return gr.update(value=log_msg), idx + 1, hf_df def skip_action(idx, lang): write_log("TARGET_SKIPPED", f"User skipped group {idx + 1}") return gr.update(value=t_text(f"Skipped group {idx + 1}.", lang)), idx + 1 # --- TRIGGER FIRST PASS --- # def trigger_first_pass(a_list, idx, b_eng_list): # if not a_list or idx >= len(a_list) or not b_eng_list: # # Return 6 updates per row (radio, html, dropdown, state_c, state_d, state_json) # return [gr.update()] * (MAX_ROWS * 6) # curr_a_eng = a_list[idx] # preds = get_model_predictions(curr_a_eng, b_eng_list) # outputs = [] # for i in range(MAX_ROWS): # if i < len(preds): # outputs.extend([ # preds[i][0], # rel_radio # preds[i][1], # conf_md # preds[i][2], # inter_dd # preds[i][3], # m_coarse_st # preds[i][4], # m_drill_st # preds[i][5], # m_conf_st # ]) # else: # outputs.extend([gr.update(), gr.update(value=""), gr.update(), "", "", ""]) # return outputs # ── EVENT WIRING ── row_outputs = [] row_inputs = [] # Notice we now unpack 12 items per row (added a_text_display) for container, a_txt, b, r, c_md, inter, j, m_co, m_dr, m_cf, m_ai_j, m_ig_j in eval_rows: row_outputs.extend([container, a_txt, b, r, c_md, inter, j, m_co, m_dr, m_cf, m_ai_j, m_ig_j]) row_inputs.extend([b, r, c_md, inter, j, m_co, m_dr, m_cf, m_ai_j, m_ig_j]) # first_pass_outputs = [] # Unpack 9 items per row # for container, b, r, c_md, inter, j, m_co, m_dr, m_cf in eval_rows: # row_outputs.extend([container, b, r, c_md, inter, j, m_co, m_dr, m_cf]) # row_inputs.extend([b, r, c_md, inter, j, m_co, m_dr, m_cf]) # first_pass_outputs.extend([r, c_md, inter, m_co, m_dr, m_cf]) # for container, b, r, c_md, inter, j, m_co, m_dr, m_cf, m_ai_j, m_ig_j in eval_rows: # row_outputs.extend([container, b, r, c_md, inter, j, m_co, m_dr, m_cf, m_ai_j, m_ig_j]) # row_inputs.extend([b, r, c_md, inter, j, m_co, m_dr, m_cf, m_ai_j, m_ig_j]) # first_pass_outputs.extend([r, c_md, inter, m_co, m_dr, m_cf]) # --- CHATBOT LOGIC --- def chat_with_ai(user_msg, history, ctx_a, ctx_b, a_list, idx, lang, user_tag, dom_a, pol_a, dom_b, pol_b): if not user_msg: yield "", history return curr_a = a_list[idx] if a_list and idx < len(a_list) else "None" system_prompt = f"You are an AI policy assistant helping an annotator understand policy documents.\nContext A: {ctx_a}\nContext B: {ctx_b}\nActive Target A: {curr_a}\nAnswer the user's query clearly and concisely based on this context." messages = [{"role": "system", "content": system_prompt}] messages.extend(history) messages.append({"role": "user", "content": user_msg}) # Append empty bot response to history for streaming history.append({"role": "user", "content": user_msg}) history.append({"role": "assistant", "content": ""}) yield "", history try: res = llm_client.chat_completion(messages=messages, max_tokens=8000, temperature=0.1, stream=True) partial_text = "" for chunk in res: token = chunk.choices[0].delta.content or "" partial_text += token # Dynamically format tags into HTML accordions as it streams history[-1]["content"] = format_streaming_thoughts(partial_text, is_streaming=True) yield "", history # Perform translation only after the stream finishes to save API calls if lang != "English": partial_text = t_text(partial_text, lang) final_formatted = format_streaming_thoughts(partial_text, is_streaming=False) history[-1]["content"] = final_formatted yield "", history # Run the background upload function log_chat_to_hf(user_tag, dom_a, pol_a, dom_b, pol_b, ctx_a, ctx_b, curr_a, user_msg, partial_text) except Exception as e: history[-1]["content"] += f"\n\nError: {str(e)}" yield "", history # Update Inputs to capture required Contexts and Domains chat_inputs = [chat_input, chatbot, ctx_a_eng_state, ctx_b_eng_state, target_a_list_state, current_index_state, lang_selector, user_tag_state, domain_a_dd, policy_a_dd, domain_b_dd, policy_b_dd] chat_submit.click(fn=chat_with_ai, inputs=chat_inputs, outputs=[chat_input, chatbot]) chat_input.submit(fn=chat_with_ai, inputs=chat_inputs, outputs=[chat_input, chatbot]) lang_selector.change( fn=handle_language_change, inputs=[ lang_selector, ctx_a_eng_state, ctx_b_eng_state, target_a_list_state, pending_tasks_state, current_index_state, user_tag_state, policy_a_dd, policy_b_dd, hf_df_state ], outputs=[ main_title, main_desc, get_started_btn, login_title, login_disclaimer, login_btn, sector_title, sector_cb, proceed_btn, app_definitions, interaction_acc, interaction_md, data_acc, back_to_sectors_btn, src_a_title, src_b_title, load_btn, bulk_title, bulk_desc, skip_btn, save_btn, workspace_info, footer_disclaimer, display_context_a, display_context_b, progress_text ] + row_outputs + [current_b_eng_list_state] ) load_btn.click( fn=load_workspace, inputs=[ domain_a_dd, policy_a_dd, domain_b_dd, policy_b_dd, target_col_dd, context_col_dd, hf_df_state, user_tag_state, lang_selector ], outputs=[ target_a_list_state, pending_tasks_state, current_index_state, ctx_a_eng_state, ctx_b_eng_state, current_b_eng_list_state, progress_text, meta_a, display_context_a, meta_b, display_context_b, status_box, workspace_box ] + row_outputs ) save_btn.click( fn=save_action, inputs=[ current_index_state, target_a_list_state, pending_tasks_state, ctx_a_eng_state, ctx_b_eng_state, current_b_eng_list_state, domain_a_dd, policy_a_dd, domain_b_dd, policy_b_dd, target_col_dd, context_col_dd, user_tag_state, session_id_state, consent_link_state, consent_follow_state, hf_df_state, lang_selector ] + row_inputs, outputs=[status_box, current_index_state, hf_df_state] ).then( fn=render_target_a, inputs=[ target_a_list_state, pending_tasks_state, current_index_state, lang_selector, user_tag_state, policy_a_dd, policy_b_dd, hf_df_state ], outputs=[progress_text] + row_outputs + [current_b_eng_list_state] ) skip_btn.click( fn=skip_action, inputs=[current_index_state, lang_selector], outputs=[status_box, current_index_state] ).then( fn=render_target_a, inputs=[ target_a_list_state, pending_tasks_state, current_index_state, lang_selector, user_tag_state, policy_a_dd, policy_b_dd, hf_df_state ], outputs=[progress_text] + row_outputs + [current_b_eng_list_state] ) # demo.launch(debug=True, ssr_mode=False, show_error=True) demo.queue(default_concurrency_limit=40).launch(debug=True, show_error=True, ssr_mode=False)