app.py

# app.py
import os
import time
import re
import requests
import phonenumbers
import pandas as pd
import urllib.parse
from bs4 import BeautifulSoup

import torch
from transformers import (
    AutoTokenizer,
    AutoModelForTokenClassification,
    AutoModelForSeq2SeqLM,
    pipeline
)

import gradio as gr
from concurrent.futures import ThreadPoolExecutor, as_completed
from email.message import EmailMessage
import smtplib
from email.mime.multipart import MIMEMultipart
from email.mime.text import MIMEText

# ============================
# CONFIG (ENV VARS recommended)
# ============================
# IMPORTANT: set these as Space "Secrets" (see README below)
API_KEY = os.environ.get("GOOGLE_API_KEY", "YOUR_GOOGLE_API_KEY")
CX      = os.environ.get("GOOGLE_CSE_ID", "YOUR_CSE_ID")
DEFAULT_COUNTRY = "Ghana"

RESULTS_PER_QUERY = int(os.environ.get("RESULTS_PER_QUERY", 4))
MAX_SCRAPE_WORKERS = int(os.environ.get("MAX_SCRAPE_WORKERS", 6))

ALLY_AI_NAME = os.environ.get("ALLY_AI_NAME", "Ally AI Assistant")
ALLY_AI_LOGO_URL_DEFAULT = os.environ.get("ALLY_AI_LOGO_URL",
                                         "https://i.ibb.co/7nZqz0H/ai-logo.png")

# Optional country maps for search bias & phone parsing
COUNTRY_TLD_MAP   = {"Ghana":"gh","Nigeria":"ng","Kenya":"ke","South Africa":"za","USA":"us","United Kingdom":"uk"}
COUNTRY_REGION_MAP= {"Ghana":"GH","Nigeria":"NG","Kenya":"KE","South Africa":"ZA","USA":"US","United Kingdom":"GB"}

# HTTP + Regex
HEADERS = {"User-Agent":"Mozilla/5.0 (X11; Linux x86_64)"}
EMAIL_REGEX = re.compile(r"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}")

# ============================
# MODELS (lightweight & CPU-friendly)
# ============================
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print("Device set to use", DEVICE)

# NER model (people/orgs/locs)
ner_model_id = "dslim/bert-base-NER"
ner_tokenizer = AutoTokenizer.from_pretrained(ner_model_id)
ner_model = AutoModelForTokenClassification.from_pretrained(ner_model_id)
ner_pipe = pipeline("ner", model=ner_model, tokenizer=ner_tokenizer, aggregation_strategy="simple",
                    device=0 if DEVICE=="cuda" else -1)

# Summarizer / anonymizer
text_model_id = "google/flan-t5-large"
text_tokenizer = AutoTokenizer.from_pretrained(text_model_id)
text_model = AutoModelForSeq2SeqLM.from_pretrained(text_model_id).to(DEVICE)

# ============================
# TAXONOMY & HELPERS
# ============================
PROFESSION_KEYWORDS = ["lawyer","therapist","doctor","counselor","social worker",
                       "advocate","psychologist","psychiatrist","consultant","nurse","hotline","gbv"]

PROBLEM_PROFESSION_MAP = {
    "rape": ["lawyer","therapist","counselor","doctor"],
    "sexual assault": ["lawyer","therapist","counselor"],
    "domestic violence": ["lawyer","social worker","therapist"],
    "abuse": ["counselor","social worker","therapist","lawyer"],
    "trauma": ["therapist","psychologist","psychiatrist"],
    "depression": ["therapist","psychologist","doctor"],
    "violence": ["lawyer","counselor","social worker"],
}

def get_region_for_country(country: str) -> str:
    return COUNTRY_REGION_MAP.get(country, "GH")

def get_tld_for_country(country: str) -> str:
    return COUNTRY_TLD_MAP.get(country, "")

def build_country_biased_query(core: str, country: str) -> str:
    tld = get_tld_for_country(country)
    suffix = f" in {country}"
    if tld:
        return f"{core}{suffix} site:.{tld} OR {country}"
    return f"{core}{suffix}"

def dedup_by_url(items):
    seen, out = set(), []
    for it in items:
        u = it.get("link") or it.get("url")
        if u and u not in seen:
            seen.add(u)
            out.append(it)
    return out

# ============================
# SEARCH & SCRAPING
# ============================
def google_search(query, num_results=5):
    if not API_KEY or not CX or "YOUR_GOOGLE_API_KEY" in API_KEY or "YOUR_CSE_ID" in CX:
        raise RuntimeError("Google API key and CSE ID must be set as environment variables.")
    url = "https://www.googleapis.com/customsearch/v1"
    params = {"q":query, "key":API_KEY, "cx":CX, "num":num_results}
    r = requests.get(url, params=params, timeout=20)
    r.raise_for_status()
    items = r.json().get("items", []) or []
    return [{"title":i.get("title",""), "link":i.get("link",""), "snippet":i.get("snippet","")} for i in items]

def extract_phones(text, region="GH"):
    phones = []
    for match in phonenumbers.PhoneNumberMatcher(text, region):
        try:
            phones.append(phonenumbers.format_number(match.number, phonenumbers.PhoneNumberFormat.INTERNATIONAL))
        except Exception:
            pass
    return list(set(phones))

def scrape_contacts(url, region="GH"):
    try:
        res = requests.get(url, headers=HEADERS, timeout=12)
        if not res.ok or not res.text:
            return {"emails": [], "phones": []}
        text = BeautifulSoup(res.text, "html.parser").get_text(separator=" ")
        text = " ".join(text.split())[:300000]
        emails = list(set(EMAIL_REGEX.findall(text)))
        phones = extract_phones(text, region)
        return {"emails": emails, "phones": phones}
    except Exception as e:
        print(f"[scrape error] {url} -> {e}")
        return {"emails": [], "phones": []}

# ============================
# NER + STORY → PROFESSIONS
# ============================
def extract_entities(text):
    if not text:
        return [],[],[]
    try:
        ner_results = ner_pipe(text)
    except Exception as e:
        print("[ner error]", e)
        return [],[],[]
    people = [e["word"] for e in ner_results if e.get("entity_group") == "PER"]
    orgs   = [e["word"] for e in ner_results if e.get("entity_group") == "ORG"]
    locs   = [e["word"] for e in ner_results if e.get("entity_group") == "LOC"]
    return list(set(people)), list(set(orgs)), list(set(locs))

def professions_from_story(story: str):
    s = (story or "").lower()
    found = set([p for p in PROFESSION_KEYWORDS if p in s])
    for prob, profs in PROBLEM_PROFESSION_MAP.items():
        if prob in s:
            found.update(profs)
    if not found:
        return ["gbv","counselor"]
    order = ["lawyer","therapist","counselor","social worker","psychologist","psychiatrist","doctor","advocate","nurse","hotline","gbv"]
    return [p for p in order if p in found]

def build_queries(story: str, country: str):
    profs = professions_from_story(story)
    cores = []
    for p in profs:
        if p == "gbv":
            cores += ["GBV support organizations", "gender based violence help"]
        else:
            cores += [f"{p} for GBV", f"{p} for sexual assault"]
    unique_cores, seen = [], set()
    for c in cores:
        if c not in seen:
            unique_cores.append(c); seen.add(c)
    return [build_country_biased_query(core, country) for core in unique_cores], profs

# ============================
# TEXT GEN: anonymize + result summary
# ============================
def anonymize_story(story: str, max_sentences: int = 2):
    if not story or not story.strip():
        return ""
    prompt = (
        "Anonymize and shorten the following personal story for contacting professionals. "
        "Remove names, exact ages, dates, locations and any identifying details. "
        f"Keep only the essential problem and the type of help requested. Output <= {max_sentences} sentences.\n\n"
        f"Story: {story}\n\nSummary:"
    )
    inputs = text_tokenizer(prompt, return_tensors="pt").to(DEVICE)
    with torch.no_grad():
        outputs = text_model.generate(**inputs, max_new_tokens=120, temperature=0.2)
    return text_tokenizer.decode(outputs[0], skip_special_tokens=True).strip()

def generate_summary(query, people, orgs, locs):
    prompt = (
        "Write a short, empathetic summary of these search results for a person seeking GBV help.\n"
        f"Query: {query}\nPeople: {', '.join(people) or '—'}\nOrgs: {', '.join(orgs) or '—'}\nLocations: {', '.join(locs) or '—'}\n\n"
        "Explain how the organizations/professionals can help in 3-4 sentences."
    )
    inputs = text_tokenizer(prompt, return_tensors="pt").to(DEVICE)
    with torch.no_grad():
        outputs = text_model.generate(**inputs, max_new_tokens=150, temperature=0.7)
    return text_tokenizer.decode(outputs[0], skip_special_tokens=True).strip()

# ============================
# MAIN PIPELINE
# ============================
def find_professionals_from_story(story, country=DEFAULT_COUNTRY, results_per_query=RESULTS_PER_QUERY):
    region = get_region_for_country(country)
    queries, profs = build_queries(story, country)

    # Search
    search_results = []
    for q in queries:
        try:
            items = google_search(q, num_results=results_per_query)
            for it in items:
                it["query"] = q
            search_results.extend(items)
        except Exception as e:
            print("[search error]", q, e)

    search_results = dedup_by_url(search_results)
    if not search_results:
        return {"summary":"No results found. Try a different country or wording.",
                "professionals":[], "queries_used":queries}

    # NER on titles/snippets
    all_people, all_orgs, all_locs = [], [], []
    for r in search_results:
        ctx = f"{r.get('title','')}. {r.get('snippet','')}"
        p,o,l = extract_entities(ctx)
        all_people += p; all_orgs += o; all_locs += l

    # Scrape contacts concurrently
    professionals = []
    with ThreadPoolExecutor(max_workers=MAX_SCRAPE_WORKERS) as ex:
        futures = {ex.submit(scrape_contacts, r["link"], region): r for r in search_results}
        for fut in as_completed(futures):
            r = futures[fut]
            contacts = {"emails": [], "phones": []}
            try:
                contacts = fut.result()
            except Exception as e:
                print("[scrape future error]", r["link"], e)
            professionals.append({
                "title": r.get("title",""),
                "url": r.get("link",""),
                "email": contacts["emails"][0] if contacts["emails"] else "Not found",
                "phone": contacts["phones"][0] if contacts["phones"] else "Not found",
                "source_query": r.get("query","")
            })

    summary = generate_summary("; ".join(queries[:3]) + (" ..." if len(queries)>3 else ""),
                               list(set(all_people)), list(set(all_orgs)), list(set(all_locs)))

    # Sort by availability of email/phone
    professionals.sort(key=lambda it: (0 if it["email"]!="Not found" else 1,
                                       0 if it["phone"]!="Not found" else 1))
    return {"summary": summary, "professionals": professionals, "queries_used": queries}

# ============================
# DRAFT (mailto + .eml)
# ============================
def build_mailto_and_eml(to_addr, subject, body, default_from="[email protected]"):
    from email.message import EmailMessage
    import time

    msg = EmailMessage()
    msg["From"] = default_from
    msg["To"] = to_addr
    msg["Subject"] = subject
    msg.set_content(body)

    # ✅ Save to a writable directory (current working dir or "tmp")
    os.makedirs("tmp", exist_ok=True)
    fname = os.path.join("tmp", f"email_draft_{int(time.time())}.eml")

    with open(fname, "wb") as f:
        f.write(msg.as_bytes())

    # Create mailto link (this part is fine)
    mailto = f"mailto:{to_addr}?subject={subject}&body={body}"

    return mailto, fname

# ============================
# SENDER (SMTP) — Ally AI branding
# ============================
def send_ally_ai_email(to_email, subject, body, user_email,
                       sender_email, sender_password,
                       ai_name=ALLY_AI_NAME,
                       logo_url=ALLY_AI_LOGO_URL_DEFAULT):
    """
    Sends an HTML email branded as Ally AI.
    to_email: recipient (organization)
    subject: subject line
    body: main message (already anonymized or full text)
    user_email: survivor's email (included for reply inside body)
    sender_email/sender_password: SMTP credentials (use Gmail App Password with Gmail)
    """
    if not to_email or to_email == "Not found":
        return "❌ No recipient email found — choose a contact with an email."

    msg = MIMEMultipart("alternative")
    msg["Subject"] = subject or "Request for support"
    msg["From"] = f"{ai_name} <{sender_email}>"
    msg["To"] = to_email

    html_content = f"""
    <html>
    <body style="font-family: Arial, sans-serif; color: #333;">
        <div style="padding: 20px; border: 1px solid #eee; border-radius: 10px; max-width: 640px; margin: auto;">
            <div style="text-align: center;">
                <img src="{logo_url}" alt="{ai_name} Logo" width="120" style="margin-bottom: 20px;" />
            </div>
            <p>{body}</p>
            <p style="margin-top:20px;">
                <b>Contact the survivor back at:</b> <a href="mailto:{user_email}">{user_email}</a>
            </p>
            <hr style="border:none;border-top:1px solid #eee;margin:24px 0;">
            <p style="font-size: 12px; color: gray; text-align: center;">
                This message was prepared with the help of <b>{ai_name}</b> — connecting survivors with help safely.
            </p>
        </div>
    </body>
    </html>
    """
    msg.attach(MIMEText(html_content, "html"))

    try:
        server = smtplib.SMTP("smtp.gmail.com", 587)
        server.starttls()
        server.login(sender_email, sender_password)  # Gmail App Password recommended
        server.sendmail(sender_email, [to_email], msg.as_string())
        server.quit()
        return f"✅ Email sent successfully to {to_email}"
    except Exception as e:
        return f"❌ Failed to send email: {str(e)}"

# ============================
# GRADIO UI
# ============================
# ------- Replace existing run_search and _on_search with these -------

def run_search(story, country):
    """
    Robust search wrapper: returns (summary, table_records, dropdown_options, anonymized_text).
    Avoids returning gr.update(...) to prevent KeyError during serialization.
    """
    try:
        out = find_professionals_from_story(story, country=country, results_per_query=RESULTS_PER_QUERY)
    except Exception as e:
        err_msg = f"Search failed: {e}"
        placeholder = ["0 — No results (search failed)"]
        return err_msg, [], placeholder, ""

    pros = out.get("professionals", []) or []

    # build table records
    try:
        records = pd.DataFrame(pros).to_dict(orient="records") if pros else []
    except Exception:
        records = []

    # build dropdown options as list of strings (guarantee at least one)
    options = []
    for i, r in enumerate(pros):
        label_contact = r.get("email") if r.get("email") and r.get("email") != "Not found" else (r.get("phone", "No contact"))
        title = r.get("title") or r.get("url") or "(no title)"
        label = f"{i} — {title} ({label_contact})"
        options.append(label)

    if not options:
        options = ["0 — No results (try a different country/query)"]

    # anonymize safely
    try:
        anon = anonymize_story(story) or "I am seeking confidential support regarding gender-based violence."
    except Exception as e:
        print("[anonymize error]", e)
        anon = "I am seeking confidential support regarding gender-based violence."

    summary = out.get("summary", "No results found.")
    return summary, records, options, anon


def _on_search(story, country):
    """
    Function wired to the search button.
    Returns exactly 5 outputs to match:
      [summary_out, results_table, dropdown_sel, anon_out, message_in]
    """
    summary, records, options, anon = run_search(story, country)

    # pre-fill message body with anonymized text (user email left empty for now)
    prefill = make_body(anon, story, True, "")

    # Return plain serializable values (not gr.update)
    # summary -> str
    # records -> list[dict] (or [])
    # options -> list[str] for dropdown (Gradio will accept it)
    # anon -> str
    # prefill -> str (message body)
    return summary, records, options, anon, prefill


def make_body(anon_text, full_story, use_anon, user_email):
    core = (anon_text or "").strip() if use_anon else (full_story or "").strip()
    # polite template with user email included in body
    lines = [
        core,
        "",
        f"Reply contact: {user_email}",
        "",
        "Thank you."
    ]
    return "\n".join([l for l in lines if l is not None])

def preview_contact(dropdown_value, df_json, subject, message_text):
    if not dropdown_value:
        return "No contact selected.", ""
    try:
        idx = int(str(dropdown_value).split(" — ")[0])
        rows = pd.DataFrame(df_json)
        contact = rows.iloc[idx].to_dict()
        recipient = contact.get("email") if contact.get("email") and contact.get("email")!="Not found" else "[no email]"
        html = f"""
        <h3>Preview</h3>
        <b>To:</b> {recipient}<br/>
        <b>Organization:</b> <a href="{contact.get('url')}" target="_blank" rel="noopener">{contact.get('title')}</a><br/>
        <b>Subject:</b> {subject}<br/>
        <hr/>
        <pre style="white-space:pre-wrap;">{message_text}</pre>
        """
        text = f"To: {recipient}\nSubject: {subject}\n\n{message_text[:600]}{'...' if len(message_text)>600 else ''}"
        return text, html
    except Exception as e:
        return f"Preview error: {e}", ""

def confirm_action(mode, dropdown_value, df_json, subject, message_text,
                   user_email, sender_email, sender_password, logo_url):
    """
    mode: "Draft only" or "Send via SMTP (Gmail)"
    """
    if not dropdown_value:
        return "❌ No contact selected.", "", None

    # locate contact
    try:
        idx = int(str(dropdown_value).split(" — ")[0])
        rows = pd.DataFrame(df_json)
        contact = rows.iloc[idx].to_dict()
    except Exception as e:
        return f"❌ Selection error: {e}", "", None

    recipient = contact.get("email")
    if mode.startswith("Send"):
        # Validate required fields
        if not recipient or recipient == "Not found":
            return "❌ This contact has no email address. Choose another contact.", "", None
        if not user_email or "@" not in user_email:
            return "❌ Please enter your email (so the organisation can contact you).", "", None
        if not sender_email or not sender_password:
            return "❌ Sender email and app password are required for SMTP sending.", "", None

        status = send_ally_ai_email(
            to_email=recipient,
            subject=subject,
            body=message_text,
            user_email=user_email,
            sender_email=sender_email,
            sender_password=sender_password,
            ai_name=ALLY_AI_NAME,
            logo_url=logo_url or ALLY_AI_LOGO_URL_DEFAULT
        )
        # also provide an .eml draft copy (optional)
        _, eml_path = build_mailto_and_eml(recipient, subject, message_text, default_from=sender_email)
        file_out = eml_path if eml_path and os.path.exists(eml_path) else None
        return status, "", file_out
    else:
        # Draft-only path
        recip_for_draft = recipient if (recipient and recipient!="Not found") else ""
        mailto, eml_path = build_mailto_and_eml(recip_for_draft, subject, message_text, default_from="[email protected]")
        html_link = f'<a href="{mailto}" target="_blank" rel="noopener">Open draft in email client</a>'
        file_out = eml_path if eml_path and os.path.exists(eml_path) else None
        return "✅ Draft created (no email sent).", html_link, file_out

with gr.Blocks() as demo:
    gr.Markdown("## Ally AI — GBV Help Finder & Email Assistant\n"
                "This tool searches local organizations, lets you select a contact, and creates an email draft or sends a branded email via SMTP.\n"
                "**Privacy tip:** Prefer anonymized summaries unless you’re comfortable sharing details.")

    with gr.Row():
        story_in   = gr.Textbox(label="Your story (free text)", lines=6, placeholder="Describe your situation and the help you want...")
        country_in = gr.Textbox(value=DEFAULT_COUNTRY, label="Country (to bias search)")

    search_btn = gr.Button("Search for professionals")
    summary_out = gr.Textbox(label="Search summary (AI)", interactive=False)
    results_table = gr.Dataframe(headers=["title","url","email","phone","source_query"], label="Search results")

    dropdown_sel = gr.Dropdown(label="Select organization (from results)", choices=[])

    with gr.Row():
        use_anon = gr.Checkbox(value=True, label="Use anonymized summary (recommended)")
        anon_out = gr.Textbox(label="Anonymized summary", lines=3)
    user_email_in = gr.Textbox(label="Your email (for the organisation to reply to you)")

    gr.Markdown("### Compose message")
    subject_in = gr.Textbox(value="Request for GBV support", label="Email subject")
    message_in = gr.Textbox(label="Message body", lines=10)

    with gr.Accordion("Sending options (for automatic sending via Ally AI SMTP)", open=False):
        mode = gr.Radio(choices=["Draft only (mailto + .eml)", "Send via SMTP (Gmail)"], value="Draft only (mailto + .eml)", label="Delivery mode")
        sender_email_in = gr.Textbox(label="Ally AI sender email (SMTP account)")
        sender_pass_in  = gr.Textbox(label="Ally AI sender app password", type="password")
        logo_url_in     = gr.Textbox(value=ALLY_AI_LOGO_URL_DEFAULT, label="Ally AI logo URL")

    with gr.Row():
        preview_btn = gr.Button("Preview")
        confirm_btn = gr.Button("Confirm (Create Draft or Send)")

    preview_text_out = gr.Textbox(label="Preview (text)", interactive=False)
    preview_html_out = gr.HTML()
    status_out = gr.Textbox(label="Status", interactive=False)
    mailto_html_out = gr.HTML()
    eml_file_out = gr.File(label="Download .eml")

    # Wire: Search
    def _on_search(story, country):
        s, records, options, anon = run_search(story, country)
        # set dropdown + anonymized text and prefill message
        prefill = make_body(anon, story, True, "")  # user email unknown yet
        return s, records, gr.update(choices=options, value=(options[0] if options else None)), anon, prefill

    search_btn.click(_on_search,
                     inputs=[story_in, country_in],
                     outputs=[summary_out, results_table, dropdown_sel, anon_out, message_in])

    # When user toggles anonymized vs full story, refresh the message body
    def _refresh_body(use_anon_flag, anon_text, story, user_email):
        return make_body(anon_text, story, use_anon_flag, user_email)

    use_anon.change(_refresh_body, inputs=[use_anon, anon_out, story_in, user_email_in], outputs=message_in)
    user_email_in.change(_refresh_body, inputs=[use_anon, anon_out, story_in, user_email_in], outputs=message_in)
    anon_out.change(_refresh_body, inputs=[use_anon, anon_out, story_in, user_email_in], outputs=message_in)
    story_in.change(_refresh_body, inputs=[use_anon, anon_out, story_in, user_email_in], outputs=message_in)

    # Preview
    preview_btn.click(preview_contact,
                      inputs=[dropdown_sel, results_table, subject_in, message_in],
                      outputs=[preview_text_out, preview_html_out])

    # Confirm (create draft or send)
    confirm_btn.click(confirm_action,
                      inputs=[mode, dropdown_sel, results_table, subject_in, message_in,
                              user_email_in, sender_email_in, sender_pass_in, logo_url_in],
                      outputs=[status_out, mailto_html_out, eml_file_out])

demo.launch(share=False)