app.py

import os
import re
import textwrap
import traceback
import pandas as pd
import gradio as gr

from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_community.vectorstores import FAISS
from langchain_core.documents import Document
from langchain_core.prompts import PromptTemplate


# =========================
# Files expected in repo root
# =========================
BASE_DIR = os.path.dirname(__file__)
CSV_FILE = os.path.join(BASE_DIR, "Customer records.csv")
RISK_PDF = os.path.join(BASE_DIR, "Risk_Policy.pdf")
INT_PDF  = os.path.join(BASE_DIR, "Interest_Rate_Policy.pdf")


# =========================
# Helpers: formatting
# =========================
def one_sentence_per_line(text: str, width: int = 110) -> str:
    if text is None:
        return ""

    def _wrap_line(line: str):
        prefix_match = re.match(r"^(\s*(?:[-*]|\d+\.|\u2022)\s+)(.*)$", line)
        if prefix_match:
            prefix, body = prefix_match.group(1), prefix_match.group(2)
            wrapped = textwrap.wrap(body, width=max(20, width - len(prefix))) or [""]
            return [prefix + wrapped[0]] + [(" " * len(prefix)) + w for w in wrapped[1:]]
        return textwrap.wrap(line, width=width) or [""]

    out_lines = []
    for raw_line in str(text).splitlines():
        line = raw_line.rstrip()
        if not line:
            out_lines.append("")
            continue
        parts = re.split(r"(?<=[.!?])\s+", line)
        for part in parts:
            part = part.strip()
            if part:
                out_lines.extend(_wrap_line(part))
    return "\n".join(out_lines)


def enforce_third_person(text: str, customer_name: str) -> str:
    if not text:
        return ""
    name = customer_name or "The customer"
    text = re.sub(r"\bYou're\b", f"{name} is", text)
    text = re.sub(r"\byou're\b", f"{name} is", text)
    text = re.sub(r"\bYou\b", name, text)
    text = re.sub(r"\byou\b", name, text)
    text = re.sub(r"\bYour\b", f"{name}'s", text)
    text = re.sub(r"\byour\b", f"{name}'s", text)
    return text


def format_customer_profile(profile: dict) -> str:
    if not profile:
        return ""
    d = dict(profile)

    preferred = ["ID", "Name", "Email", "Credit_Score", "Nationality", "Account_Status", "PR_Status"]

    nat = str(d.get("Nationality", "")).strip().lower()
    if nat == "singaporean":
        d.pop("PR_Status", None)

    lines = []
    for k in preferred:
        if k in d:
            lines.append(f"{k}: {d.get(k)}")
    for k in sorted(d.keys()):
        if k not in preferred:
            lines.append(f"{k}: {d.get(k)}")
    return "\n".join(lines)


# =========================
# Load CSV + build mock systems
# =========================
def load_customer_csv(csv_path: str) -> pd.DataFrame:
    df_all = pd.read_csv(csv_path)
    df_all.columns = [c.strip() for c in df_all.columns]

    df_all["ID"] = df_all["ID"].astype(str).str.strip()
    df_all["Credit Score"] = pd.to_numeric(df_all["Credit Score"], errors="coerce").astype("Int64")
    df_all["Account Status"] = df_all["Account Status"].astype(str).str.strip().str.lower()
    df_all["Nationality"] = df_all["Nationality"].astype(str).str.strip()

    def _parse_pr(x):
        if pd.isna(x):
            return pd.NA
        s = str(x).strip().lower()
        if s in ("true", "t", "yes", "y", "1"):
            return True
        if s in ("false", "f", "no", "n", "0"):
            return False
        return pd.NA

    if "PR Status" in df_all.columns:
        df_all["PR_Status"] = df_all["PR Status"].apply(_parse_pr)
    else:
        df_all["PR_Status"] = pd.NA

    return df_all


def build_mock_systems(df_all: pd.DataFrame):
    df_credit = df_all[["ID", "Name", "Email", "Credit Score"]].copy()
    df_credit = df_credit.rename(columns={"Credit Score": "Credit_Score"})

    df_account = df_all[["ID", "Name", "Nationality", "Account Status"]].copy()
    df_account = df_account.rename(columns={"Account Status": "Account_Status"})

    df_pr = df_all.dropna(subset=["PR_Status"]).copy()
    df_gov = df_pr[["ID", "Name", "PR_Status"]].copy()

    return df_credit, df_account, df_gov


def get_customer_profile(customer_id: str):
    customer_id = str(customer_id).strip()
    credit_rec = DF_CREDIT[DF_CREDIT["ID"].astype(str) == customer_id]
    if credit_rec.empty:
        return None

    name = credit_rec.iloc[0]["Name"]
    email = credit_rec.iloc[0]["Email"]
    credit_score = int(credit_rec.iloc[0]["Credit_Score"])

    acct_rec = DF_ACCOUNT[DF_ACCOUNT["ID"].astype(str) == customer_id]
    nationality = acct_rec.iloc[0]["Nationality"] if not acct_rec.empty else None
    account_status = acct_rec.iloc[0]["Account_Status"] if not acct_rec.empty else None

    pr_status = None
    if nationality and str(nationality).strip().lower() == "non-singaporean":
        gov_rec = DF_GOV[DF_GOV["ID"].astype(str) == customer_id]
        pr_status = bool(gov_rec.iloc[0]["PR_Status"]) if not gov_rec.empty else None

    return {
        "ID": customer_id,
        "Name": name,
        "Email": email,
        "Credit_Score": credit_score,
        "Nationality": nationality,
        "Account_Status": account_status,
        "PR_Status": pr_status,
    }


# =========================
# PDF ingest + parse policies
# =========================
def extract_pdf_text(pdf_path: str) -> str:
    from pypdf import PdfReader
    reader = PdfReader(pdf_path)
    pages = []
    for p in reader.pages:
        pages.append(p.extract_text() or "")
    return "\n".join(pages)


def parse_policies(risk_text: str, interest_text: str):
    # Interest rates
    rate_matches = re.findall(
        r"\b(Low|Medium|High)\b\s+([0-9]+\.[0-9]+)\s*%?",
        interest_text,
        flags=re.IGNORECASE
    )
    interest_rates = {k.capitalize(): float(v) for k, v in rate_matches}

    # Risk table rows
    risk_rows = re.findall(
        r"(\d{3})\s*(?:-|–|—)?\s*(\d{3})\s+(Delinquent|Closed|Good-standing)\s+(High|Medium|Low)",
        risk_text,
        flags=re.IGNORECASE,
    )

    risk_mapping = {}
    for lo, hi, status, risk in risk_rows:
        band = (int(lo), int(hi))
        risk_mapping[(band, status.strip().lower())] = risk.capitalize()

    return risk_mapping, interest_rates


def _score_band(score: int):
    if 300 <= score <= 674: return (300, 674)
    if 675 <= score <= 749: return (675, 749)
    if 750 <= score <= 850: return (750, 850)
    if score < 300: return (300, 674)
    return (750, 850)


def determine_overall_risk(score: int, account_status: str) -> str:
    band = _score_band(score)
    status = str(account_status).strip().lower()
    return RISK_MAPPING.get((band, status), "High")


def determine_interest_rate(overall_risk: str) -> float:
    return float(INTEREST_RATES[overall_risk])


def is_non_singaporean_no_pr(customer_id: str) -> bool:
    cid = str(customer_id).strip()
    nat_row = DF_ACCOUNT[DF_ACCOUNT["ID"].astype(str).str.strip() == cid]
    nationality = nat_row.iloc[0]["Nationality"] if not nat_row.empty else None

    pr_row = DF_GOV[DF_GOV["ID"].astype(str).str.strip() == cid]
    pr_status = bool(pr_row.iloc[0]["PR_Status"]) if not pr_row.empty else False

    return (str(nationality).strip().lower() != "singaporean") and (pr_status is False)


def apply_mandatory_exception_to_report(report_text: str, customer_id: str) -> str:
    text = "" if report_text is None else str(report_text)

    if not is_non_singaporean_no_pr(customer_id):
        return text

    text = re.sub(
        r"(Loan\s+Approval\s+Decision:\s*\*\*?)\s*Approve\b",
        r"\1 Reject",
        text,
        flags=re.IGNORECASE
    )
    text = re.sub(
        r"(Loan\s+Approval:\s*\*\*?)\s*Approve\b",
        r"\1 Reject",
        text,
        flags=re.IGNORECASE
    )

    if not re.search(r"not\s+recommend|reject", text, flags=re.IGNORECASE):
        text += "\n\nConclusion: Not recommend although risk is low, because Non-Singaporean and PR status is false."
    elif not re.search(r"Non[-\s]?Singaporean.*PR", text, flags=re.IGNORECASE | re.DOTALL):
        text += "\n\nConclusion: Not recommend although risk is low, because Non-Singaporean and PR status is false."

    return text


# =========================
# Unstructured resolver (Colab-style)
# =========================
def resolve_customer_id(unstructured_text: str):
    s = (unstructured_text or "").strip()
    if not s:
        return None, "❌ Please enter Applicant Name or ID."

    # 1) Extract an ID from any sentence
    m = re.search(r"\b(\d{3,})\b", s)
    if m:
        cid = m.group(1)
        prof = get_customer_profile(cid)
        if prof:
            return cid, f"✅ Found ID {cid}: {prof['Name']}"
        return None, f"❌ No such customer ID: {cid}"

    # 2) Otherwise treat as name search (contains)
    results = DF_CREDIT[DF_CREDIT["Name"].astype(str).str.contains(s, case=False, na=False)]
    if results.empty:
        return None, f"❌ No such customer: '{s}'"

    if len(results) == 1:
        cid = str(results.iloc[0]["ID"])
        nm = str(results.iloc[0]["Name"])
        return cid, f"✅ Found Name '{nm}' -> ID {cid}"

    # Multiple matches: no dropdown, just tell user to type ID
    opts = ", ".join([f"{r['Name']} (ID {r['ID']})" for _, r in results.iterrows()])
    return None, f"⚠️ Multiple customers match '{s}'. Please enter ID. Matches: {opts}"


# =========================
# Prompts / Chains
# =========================
QA_PROMPT = PromptTemplate(
    input_variables=["customer_data", "policy_rules", "question"],
    template="""
You are a helpful banking assistant. Answer the user's question based strictly on the provided Customer Data and Policy Rules.

CUSTOMER DATA:
{customer_data}

POLICY RULES:
{policy_rules}

USER QUESTION:
{question}

ANSWER:
"""
)

ADVICE_PROMPT = PromptTemplate(
    input_variables=["customer_data", "policy_rules", "question"],
    template="""You are a helpful loan officer assistant.

Write in THIRD PERSON about the customer. Always use the customer's Name and possessive.
Never address the reader as 'you' or 'your'.

You must provide ADVICE/RECOMMENDATION (not just restating risk and rate).
Use ONLY the provided Customer Data and Policy Rules.

REQUIREMENTS:
- Provide 3-5 actionable advice points (short sentences).
- Include a clear final recommendation: APPROVE or NOT RECOMMEND / REJECT.
- If customer is Non-Singaporean and PR_Status is False, you MUST recommend NOT RECOMMEND / REJECT regardless of risk level.
- Keep it concise.

CUSTOMER DATA:
{customer_data}

POLICY RULES:
{policy_rules}

USER QUESTION:
{question}

ANSWER:
"""
)

REPORT_PROMPT = PromptTemplate(
    input_variables=["customer_data", "policy_rules"],
    template="""
You are a senior loan officer. Generate a comprehensive loan assessment report based on the provided customer data and banking policies.
Analyze the customer's profile, determine overall risk, calculate interest rate, and provide a clear recommendation.
Follow any mandatory exceptions.

CUSTOMER DATA:
{customer_data}

POLICY RULES:
{policy_rules}

REPORT:
"""
)


# =========================
# Global init (Spaces safe)
# =========================
RETRIEVER = None
POLICY_FULL = ""
LLM = None

INIT_ERROR = None

try:
    # Check files
    missing = [p for p in [CSV_FILE, RISK_PDF, INT_PDF] if not os.path.exists(p)]
    if missing:
        raise RuntimeError("Missing required files in repo root:\n" + "\n".join(missing))

    # Clean key (fixes illegal header newline)
    api_key = os.getenv("OPENAI_API_KEY", "")
    api_key = api_key.strip()
    os.environ["OPENAI_API_KEY"] = api_key
    if not api_key:
        raise RuntimeError("Missing OPENAI_API_KEY. Set it in Space Secrets.")

    # Load CSV
    DF_ALL = load_customer_csv(CSV_FILE)
    DF_CREDIT, DF_ACCOUNT, DF_GOV = build_mock_systems(DF_ALL)

    # Load PDFs
    RISK_TEXT = extract_pdf_text(RISK_PDF)
    INT_TEXT  = extract_pdf_text(INT_PDF)
    POLICY_FULL = RISK_TEXT + "\n\n" + INT_TEXT

    # Parse policies
    RISK_MAPPING, INTEREST_RATES = parse_policies(RISK_TEXT, INT_TEXT)

    # LLM
    model_name = os.getenv("OPENAI_MODEL", "gpt-4o")
    LLM = ChatOpenAI(model=model_name, temperature=0, openai_api_key=api_key)

    # Chains (LCEL)
    QA_CHAIN = QA_PROMPT | LLM
    ADVICE_CHAIN = ADVICE_PROMPT | LLM
    REPORT_CHAIN = REPORT_PROMPT | LLM

except Exception as e:
    INIT_ERROR = str(e)


def build_retriever_if_needed():
    global RETRIEVER
    if RETRIEVER is not None:
        return RETRIEVER

    api_key = os.getenv("OPENAI_API_KEY", "").strip()
    if not api_key:
        return None

    docs = [
        Document(page_content=RISK_TEXT, metadata={"source": "Risk_Policy.pdf"}),
        Document(page_content=INT_TEXT,  metadata={"source": "Interest_Rate_Policy.pdf"}),
    ]
    embeddings = OpenAIEmbeddings(openai_api_key=api_key)
    db = FAISS.from_documents(docs, embeddings)
    RETRIEVER = db.as_retriever()
    return RETRIEVER


def get_policy_context(use_rag: bool) -> str:
    if not use_rag:
        return POLICY_FULL

    try:
        retriever = build_retriever_if_needed()
        if retriever is None:
            return POLICY_FULL + "\n\n[Note] RAG unavailable (missing API key). Using full policy text."

        docs = retriever.invoke("risk level interest rate PR status")
        ctx = "\n\n".join([d.page_content for d in docs]).strip()
        return ctx if ctx else POLICY_FULL
    except Exception as e:
        return POLICY_FULL + f"\n\n[Note] RAG failed, using full policy text: {e}"


# =========================
# Main Run (Colab-style)
# =========================
def run_action(user_input: str, action: str, use_rag: bool):
    if INIT_ERROR:
        return f"❌ Initialization error:\n\n{INIT_ERROR}"

    try:
        customer_id, msg = resolve_customer_id(user_input)
        if not customer_id:
            return msg

        profile = get_customer_profile(customer_id)
        if not profile:
            return f"❌ No such customer ID: {customer_id}"

        profile_text = format_customer_profile(profile)
        policy_context = get_policy_context(use_rag)

        # Deterministic summary (optional but helps demo)
        overall_risk = determine_overall_risk(profile["Credit_Score"], profile["Account_Status"])
        rate = determine_interest_rate(overall_risk)
        must_reject = is_non_singaporean_no_pr(customer_id)

        det = [
            msg,
            "",
            "Deterministic (Policy-based):",
            f"- Overall risk: {overall_risk}",
            f"- Interest rate: {rate:.3f}%",
        ]
        if must_reject:
            det.append("- Mandatory exception: NOT RECOMMEND / REJECT (Non-Singaporean without PR).")
        det_text = "\n".join(det)

        if action == "1) Check Risk & Interest":
            question = f"What are the risk level and applicable interest rate for the customer {customer_id}?"
            resp = QA_CHAIN.invoke({
                "customer_data": profile_text,
                "policy_rules": policy_context,
                "question": question
            })
            return det_text + "\n\nAI Output:\n" + one_sentence_per_line(resp.content)

        if action == "2) Advice / Recommendation":
            question = f"What interest rate advice can be recommended for customer with Id {customer_id}?"
            resp = ADVICE_CHAIN.invoke({
                "customer_data": profile_text,
                "policy_rules": policy_context,
                "question": question
            })
            return det_text + "\n\nAI Output:\n" + one_sentence_per_line(enforce_third_person(resp.content, profile.get("Name", "Customer")))

        # 3) FULL report
        customer_data_with_exception = profile_text + (
            "\n\nMANDATORY EXCEPTION (must follow): Non-Singaporean with PR_Status = False => NOT RECOMMENDED / REJECTED."
            if must_reject else ""
        )
        full_report = REPORT_CHAIN.invoke({
            "customer_data": customer_data_with_exception,
            "policy_rules": policy_context
        })
        final_text = apply_mandatory_exception_to_report(full_report.content, customer_id)
        return det_text + "\n\nFull Report:\n" + one_sentence_per_line(final_text)

    except Exception:
        return "❌ Run failed:\n\n" + traceback.format_exc()


# =========================
# Gradio UI (NO dropdown)
# =========================
with gr.Blocks(title="Bank Loan Officer System") as demo:
    gr.Markdown("# 🏦 Bank Loan Officer System (Unstructured Input Search)")
    gr.Markdown("Type Applicant **Name / ID / sentence**. The system resolves and responds like your Colab notebook.")

    if INIT_ERROR:
        gr.Markdown(f"## ❌ Initialization error\n\n```\n{INIT_ERROR}\n```")

    user_input = gr.Textbox(
        label="Applicant Name or ID (unstructured)",
        placeholder="e.g. 3333 OR Hilda OR 'please check loan for 3333'"
    )

    action = gr.Radio(
        label="Action",
        choices=[
            "1) Check Risk & Interest",
            "2) Advice / Recommendation",
            "3) FULL Formal Loan Report"
        ],
        value="1) Check Risk & Interest"
    )

    use_rag = gr.Checkbox(
        label="Use RAG (FAISS embeddings). If it fails, auto fallback.",
        value=False
    )

    run_btn = gr.Button("🚀 Run")
    output = gr.Textbox(label="Output", lines=28)

    run_btn.click(fn=run_action, inputs=[user_input, action, use_rag], outputs=[output])
    user_input.submit(fn=run_action, inputs=[user_input, action, use_rag], outputs=[output])


# =========================
# HF Spaces launch (the add-on at bottom)
# =========================
PORT = int(os.environ.get("PORT", 7860))
demo.queue().launch(
    server_name="0.0.0.0",
    server_port=PORT,
    ssr_mode=False,
    prevent_thread_lock=False
)