app.py

"""
RegTech BR — Hugging Face Space
================================
Compliance analyzer for Brazilian crypto asset regulation.
Uses FAISS RAG + Claude Sonnet 4.6 to produce structured risk assessments.

Setup:
  1. Upload faiss_index.bin, embeddings.npy, chunks_meta.jsonl to data/index/
  2. Add ANTHROPIC_API_KEY as a Space Secret
  3. Deploy
"""

import os, json, re, time, unicodedata
from pathlib import Path
import numpy as np
import gradio as gr
import requests

# ── Paths ──────────────────────────────────────────────────────────
INDEX_DIR = Path("data/index")

# ── Load index at startup ──────────────────────────────────────────
print("Loading RegTech BR index...")

import faiss
from sentence_transformers import SentenceTransformer

CHUNKS: list[dict] = []
with open(INDEX_DIR / "chunks_meta.jsonl", encoding="utf-8") as f:
    for line in f:
        line = line.strip()
        if line:
            CHUNKS.append(json.loads(line))

EMBEDDINGS = np.load(INDEX_DIR / "embeddings.npy")
INDEX      = faiss.read_index(str(INDEX_DIR / "faiss_index.bin"))
MODEL      = SentenceTransformer("paraphrase-multilingual-MiniLM-L12-v2")

print(f"Index ready — {len(CHUNKS)} chunks, {INDEX.ntotal} vectors")

# ── Normalization & routing ────────────────────────────────────────
def normalize(text: str) -> str:
    text = unicodedata.normalize("NFD", text or "")
    text = "".join(c for c in text if unicodedata.category(c) != "Mn")
    text = text.lower()
    text = re.sub(r"[^a-z0-9]+", " ", text)
    return re.sub(r"\s+", " ", text).strip()

AUTHORITY_KW = {
    "BCB":  ["banco central","bcb","psav","ativo virtual","ativos virtuais",
              "criptoativo","autorizacao","cambio","segregacao","patrimonial",
              "ativos dos clientes","circular","resolucao bcb"],
    "CVM":  ["cvm","valores mobiliarios","valor mobiliario","token","tokens",
              "oferta publica","dividendos","direito de voto","captar","rwa","parecer"],
    "COAF": ["coaf","pessoa exposta politicamente","pep","kyc","lavagem",
              "terrorismo","pld","ftp","anonimo","anonima","identificacao do cliente"],
}

SOURCE_KW = {
    "lei_14478":        ["lei 14 478","lei 14478","marco legal","lei dos criptoativos"],
    "decreto_11563":    ["decreto 11 563","decreto 11563"],
    "bcb_circular_3978":["circular 3978","pld","ftp","lavagem","kyc","anonimo"],
    "bcb_in701":        ["instrucao normativa 701","certificacao tecnica","segregacao"],
    "bcb_res548":       ["resolucao 548","autorizacao psav"],
    "cvm":              ["cvm","valores mobiliarios","token","dividendos"],
    "coaf":             ["coaf","pep","pessoa exposta politicamente"],
}

def detect_route(query: str) -> dict:
    q = normalize(query)
    route = {"authority_filters":[], "source_id_contains":[], "query_expansion":[]}

    if any(normalize(k) in q for k in SOURCE_KW["lei_14478"]):
        route["source_id_contains"].append("lei_14478")
        route["query_expansion"].extend(["Lei 14.478 de 2022","marco legal criptoativos"])

    if any(normalize(k) in q for k in SOURCE_KW["decreto_11563"]):
        route["source_id_contains"].append("decreto_11563")
        route["query_expansion"].extend(["Decreto 11.563 de 2023","Banco Central"])

    hits = {a: sum(1 for k in kws if normalize(k) in q)
            for a, kws in AUTHORITY_KW.items()}
    max_h = max(hits.values()) if hits else 0
    if max_h:
        route["authority_filters"] = [a for a,h in hits.items() if h==max_h or h>=2]

    for key, kws in SOURCE_KW.items():
        if any(normalize(k) in q for k in kws):
            route["source_id_contains"].append(key)

    for k in ["authority_filters","source_id_contains","query_expansion"]:
        route[k] = list(dict.fromkeys(route[k]))
    return route

def route_boost(chunk: dict, route: dict) -> float:
    boost = 0.0
    sid = normalize(chunk.get("source_id",""))
    for token in route.get("source_id_contains",[]):
        if normalize(token) in sid:
            boost += 0.35
    if chunk.get("authority") in route.get("authority_filters",[]):
        boost += 0.08
    return min(boost, 0.55)

def lexical_boost(query_norm: str, chunk: dict) -> float:
    parts = " ".join([chunk.get("source_id",""), chunk.get("authority",""),
                      " ".join(chunk.get("tags",[])), chunk.get("text","")])
    chunk_norm = normalize(parts)
    terms = [t for t in query_norm.split() if len(t) >= 4]
    if not terms:
        return 0.0
    hits = sum(1 for t in set(terms) if t in chunk_norm)
    return min(0.12, hits / max(6, len(set(terms))) * 0.12)

def retrieve(query: str, top_k: int = 6) -> list[dict]:
    route = detect_route(query)
    expanded = query + "\n" + "\n".join(route.get("query_expansion",[]))
    q_vec = MODEL.encode([expanded], normalize_embeddings=True).astype(np.float32)
    scores, indices = INDEX.search(q_vec, min(len(CHUNKS), top_k * 20))
    q_norm = normalize(expanded)
    ranked = []
    for score, idx in zip(scores[0], indices[0]):
        if idx < 0 or float(score) < 0.20:
            continue
        chunk = CHUNKS[int(idx)].copy()
        chunk["_score"]  = float(score)
        chunk["_final"]  = float(score) + lexical_boost(q_norm, chunk) + route_boost(chunk, route)
        ranked.append(chunk)
    ranked.sort(key=lambda r: r["_final"], reverse=True)
    seen, unique = set(), []
    for r in ranked:
        cid = r.get("chunk_id")
        if cid not in seen:
            seen.add(cid)
            unique.append(r)
    return unique[:top_k]

def format_context(results: list[dict]) -> str:
    lines = []
    for i, r in enumerate(results, 1):
        art  = f" — {r['article_hint']}" if r.get("article_hint") else ""
        norm = f" [{r['normative_reference_hint']}]" if r.get("normative_reference_hint") else ""
        lines.append(
            f"[SOURCE {i}] {r.get('source_label','')}{art}{norm}\n"
            f"Authority: {r.get('authority','?')} | Score: {r['_final']:.3f}\n"
            f"{str(r.get('text',''))[:700]}..."
        )
    return "\n\n---\n\n".join(lines)

# ── LLM ───────────────────────────────────────────────────────────
SYSTEM = """You are RegTech BR, a specialist AI in Brazilian crypto asset regulation.
Analyze the compliance query and produce a structured JSON assessment.
Respond ONLY with valid JSON — no markdown fences.

Schema:
{
  "risk_level": "LOW | MEDIUM | HIGH | UNCLEAR",
  "compliance_status": "COMPLIANT | NON-COMPLIANT | REQUIRES_REVIEW | INSUFFICIENT_INFO",
  "applicable_regulations": ["list"],
  "relevant_articles": ["list"],
  "finding": "2-5 sentence assessment",
  "corrective_action": "specific steps or 'No action required'",
  "confidence": "HIGH | MEDIUM | LOW",
  "authority": "BCB | CVM | COAF | mixed | federal"
}"""

def call_claude(query: str, context: str) -> dict | None:
    api_key = os.environ.get("ANTHROPIC_API_KEY", "")
    if not api_key:
        return None
    prompt = (f"COMPLIANCE QUERY:\n{query}\n\n"
              f"REGULATORY CONTEXT:\n\n{context}\n\n"
              f"Produce a structured compliance assessment.")
    try:
        r = requests.post(
            "https://api.anthropic.com/v1/messages",
            headers={"Content-Type":"application/json",
                     "x-api-key": api_key,
                     "anthropic-version":"2023-06-01"},
            json={"model":"claude-sonnet-4-6","max_tokens":1200,
                  "system":SYSTEM,"messages":[{"role":"user","content":prompt}]},
            timeout=90,
        )
        r.raise_for_status()
        raw = "".join(b.get("text","") for b in r.json().get("content",[])
                      if b.get("type")=="text")
        raw = re.sub(r"^```(?:json)?|```$","",raw.strip(),flags=re.IGNORECASE).strip()
        start, end = raw.find("{"), raw.rfind("}")
        if start >= 0 and end > start:
            return json.loads(raw[start:end+1])
    except Exception as e:
        print(f"Claude error: {e}")
    return None

# ── Risk badge ─────────────────────────────────────────────────────
RISK_COLOR = {"HIGH":"#dc2626","MEDIUM":"#d97706","LOW":"#16a34a","UNCLEAR":"#6b7280"}
STATUS_ICON = {"NON-COMPLIANT":"⛔","COMPLIANT":"✅","REQUIRES_REVIEW":"⚠️","INSUFFICIENT_INFO":"❓"}

def render_report(report: dict, query: str, results: list[dict]) -> str:
    risk   = report.get("risk_level","UNCLEAR")
    status = report.get("compliance_status","INSUFFICIENT_INFO")
    color  = RISK_COLOR.get(risk,"#6b7280")
    icon   = STATUS_ICON.get(status,"❓")
    conf   = report.get("confidence","LOW")
    auth   = report.get("authority","?")

    regs = "".join(f"<li>{r}</li>" for r in report.get("applicable_regulations",[]))
    arts = "".join(f"<li>{a}</li>" for a in report.get("relevant_articles",[]))
    srcs = "".join(
        f'<span style="background:#1e3a5f;color:#93c5fd;padding:2px 8px;'
        f'border-radius:4px;font-size:0.75rem;margin:2px;display:inline-block">'
        f'{r.get("source_id","")}</span>'
        for r in results
    )

    return f"""
<div style="font-family:'IBM Plex Mono',monospace;background:#0f172a;color:#e2e8f0;
            padding:1.5rem;border-radius:12px;border:1px solid #1e3a5f;line-height:1.6">

  <div style="display:flex;gap:1rem;align-items:center;margin-bottom:1.2rem;flex-wrap:wrap">
    <span style="background:{color};color:#fff;padding:4px 14px;border-radius:6px;
                 font-weight:700;font-size:0.9rem;letter-spacing:0.05em">
      {risk} RISK
    </span>
    <span style="background:#1e293b;color:#e2e8f0;padding:4px 14px;border-radius:6px;
                 font-size:0.9rem;border:1px solid #334155">
      {icon} {status}
    </span>
    <span style="color:#64748b;font-size:0.8rem">
      confidence: {conf} · authority: {auth}
    </span>
  </div>

  <div style="background:#1e293b;border-radius:8px;padding:1rem;margin-bottom:1rem;
              border-left:3px solid {color}">
    <div style="color:#94a3b8;font-size:0.75rem;text-transform:uppercase;
                letter-spacing:0.1em;margin-bottom:0.5rem">Finding</div>
    <div style="color:#e2e8f0;font-size:0.88rem">{report.get('finding','—')}</div>
  </div>

  <div style="background:#1e293b;border-radius:8px;padding:1rem;margin-bottom:1rem">
    <div style="color:#94a3b8;font-size:0.75rem;text-transform:uppercase;
                letter-spacing:0.1em;margin-bottom:0.5rem">Corrective Action</div>
    <div style="color:#e2e8f0;font-size:0.88rem">{report.get('corrective_action','—')}</div>
  </div>

  <div style="display:grid;grid-template-columns:1fr 1fr;gap:1rem;margin-bottom:1rem">
    <div style="background:#1e293b;border-radius:8px;padding:1rem">
      <div style="color:#94a3b8;font-size:0.75rem;text-transform:uppercase;
                  letter-spacing:0.1em;margin-bottom:0.5rem">Applicable Regulations</div>
      <ul style="margin:0;padding-left:1.2rem;color:#93c5fd;font-size:0.8rem">{regs}</ul>
    </div>
    <div style="background:#1e293b;border-radius:8px;padding:1rem">
      <div style="color:#94a3b8;font-size:0.75rem;text-transform:uppercase;
                  letter-spacing:0.1em;margin-bottom:0.5rem">Relevant Articles</div>
      <ul style="margin:0;padding-left:1.2rem;color:#86efac;font-size:0.8rem">{arts}</ul>
    </div>
  </div>

  <div style="margin-top:0.8rem">
    <div style="color:#64748b;font-size:0.7rem;text-transform:uppercase;
                letter-spacing:0.1em;margin-bottom:0.4rem">Sources retrieved</div>
    {srcs}
  </div>
</div>"""

def render_error(msg: str) -> str:
    return f"""
<div style="background:#1c1917;color:#fca5a5;padding:1.2rem;border-radius:8px;
            border:1px solid #7f1d1d;font-family:monospace">
  ⚠️ {msg}
</div>"""

# ── Gradio interface ───────────────────────────────────────────────
EXAMPLES = [
    "Nossa plataforma permite compra e venda de criptoativos sem autorização formal do Banco Central.",
    "Realizamos KYC apenas para transações acima de R$100.000. Transações menores são anônimas.",
    "Nosso token REV distribui dividendos e dá direito de voto. Será ofertado publicamente por R$30M.",
    "Nossa exchange não realiza segregação patrimonial dos ativos dos clientes.",
    "Não possuímos diretor responsável formalmente designado para PLD/FTP.",
    "Qual é o marco legal dos criptoativos no Brasil segundo a Lei 14.478 de 2022?",
    "Qual órgão foi designado pelo Decreto 11.563/2023 para regular as PSAVs?",
]

def analyze(query: str) -> tuple[str, str]:
    if not query or not query.strip():
        return render_error("Please enter a compliance query."), ""

    results = retrieve(query.strip())
    if not results:
        return render_error("No relevant regulatory chunks found. Try rephrasing your query."), ""

    context = format_context(results)
    report  = call_claude(query, context)

    if not report:
        return render_error(
            "Could not reach Claude API. Check that ANTHROPIC_API_KEY is set as a Space Secret."
        ), context

    return render_report(report, query, results), context

# ── Layout ─────────────────────────────────────────────────────────
CSS = """
@import url('https://fonts.googleapis.com/css2?family=IBM+Plex+Mono:wght@400;500;700&family=IBM+Plex+Sans:wght@300;400;600&display=swap');

body { background: #0a0f1e !important; }

.gradio-container {
    background: #0a0f1e !important;
    font-family: 'IBM Plex Sans', sans-serif !important;
    max-width: 900px !important;
    margin: 0 auto !important;
}

#header {
    text-align: center;
    padding: 2rem 0 1rem;
    border-bottom: 1px solid #1e3a5f;
    margin-bottom: 1.5rem;
}

#header h1 {
    font-family: 'IBM Plex Mono', monospace;
    font-size: 1.8rem;
    color: #38bdf8;
    letter-spacing: -0.02em;
    margin: 0 0 0.3rem;
}

#header p {
    color: #64748b;
    font-size: 0.85rem;
    margin: 0;
}

.query-box textarea {
    background: #0f172a !important;
    border: 1px solid #1e3a5f !important;
    color: #e2e8f0 !important;
    font-family: 'IBM Plex Mono', monospace !important;
    font-size: 0.9rem !important;
    border-radius: 8px !important;
}

.query-box textarea:focus {
    border-color: #38bdf8 !important;
    box-shadow: 0 0 0 2px rgba(56,189,248,0.15) !important;
}

.analyze-btn {
    background: #0369a1 !important;
    color: #fff !important;
    font-family: 'IBM Plex Mono', monospace !important;
    font-weight: 600 !important;
    letter-spacing: 0.05em !important;
    border: none !important;
    border-radius: 8px !important;
    height: 44px !important;
    font-size: 0.85rem !important;
    transition: background 0.2s !important;
}

.analyze-btn:hover {
    background: #0284c7 !important;
}

.context-box textarea {
    background: #0f172a !important;
    border: 1px solid #1e293b !important;
    color: #64748b !important;
    font-family: 'IBM Plex Mono', monospace !important;
    font-size: 0.75rem !important;
}

label { color: #94a3b8 !important; font-size: 0.8rem !important; }

.svelte-1gfkn6j { background: #0a0f1e !important; }
"""

with gr.Blocks(css=CSS, title="RegTech BR") as demo:
    gr.HTML("""
    <div id="header">
      <h1>⚖ RegTech BR</h1>
      <p>Brazilian Crypto Asset Compliance Analyzer · BCB · CVM · COAF · Lei 14.478/2022</p>
      <p style="color:#1e3a5f;font-size:0.75rem;margin-top:0.3rem">
        103 regulatory chunks · RAG + Claude Sonnet 4.6 · BLI Legal Tech Hackathon 2
      </p>
    </div>
    """)

    with gr.Row():
        with gr.Column(scale=4):
            query_box = gr.Textbox(
                label="Compliance query or document excerpt",
                placeholder="Describe your policy, product, or compliance question in Portuguese or English...",
                lines=4,
                elem_classes=["query-box"],
            )
        with gr.Column(scale=1, min_width=120):
            analyze_btn = gr.Button("Analyze →", elem_classes=["analyze-btn"])

    gr.Examples(
        examples=[[e] for e in EXAMPLES],
        inputs=[query_box],
        label="Example queries",
    )

    report_html = gr.HTML(label="Compliance Assessment")

    with gr.Accordion("Retrieved regulatory context", open=False):
        context_box = gr.Textbox(
            label="Raw chunks retrieved by RAG",
            lines=10,
            interactive=False,
            elem_classes=["context-box"],
        )

    gr.HTML("""
    <div style="text-align:center;color:#334155;font-size:0.72rem;
                padding:1.5rem 0 0.5rem;font-family:'IBM Plex Mono',monospace">
      ⚠ Experimental pipeline. Not legal advice. Results require review by qualified professionals.<br>
      RegTech BR · Fernando Rodrigues · Kaggle: fernandosr85
    </div>
    """)

    analyze_btn.click(fn=analyze, inputs=[query_box], outputs=[report_html, context_box])
    query_box.submit(fn=analyze, inputs=[query_box], outputs=[report_html, context_box])

if __name__ == "__main__":
    demo.launch()