Spaces:

gaurv007
/

ClauseGuard

Sleeping

App Files Files Community

ClauseGuard / app.py

gaurv007

fix(v4.3.1): app.py — Run 4 delta fixes (A-E)

ccf342b verified 12 days ago

raw

history blame contribute delete

101 kB

	"""
	ClauseGuard — World's Best Legal Contract Analysis Tool (v4.3)
	═══════════════════════════════════════════════════════════════
	PERF v4.3:
	• PERF: Upgraded embedder to BAAI/bge-small-en-v1.5 (+21% retrieval accuracy)
	• PERF: Batched clause classification (single forward pass, batch_size=8)
	• PERF: ONNX INT8 quantized model support (2-4x faster on CPU)
	• PERF: torch.set_num_threads(2) to prevent CPU thrashing
	• NEW: ml/export_onnx_v2.py — full merge→ONNX→quantize pipeline

	Fixes in v4.2:
	• FIX: NLI now uses CrossEncoder.predict() — contradictions actually work
	• FIX: BoundedCache uses threading.RLock — no more race conditions
	• FIX: Pre-compiled ALL regex patterns at module level (perf)
	• FIX: Added missing regex labels to RISK_MAP/DESC_MAP
	• FIX: Extension risk formula matches backend
	• FIX: Extension API_BASE URL corrected
	• FIX: API CORS localhost requires explicit opt-in

	Fixes in v4.1:
	• FIX: Bounded LRU caches (chunk_cache, prediction_cache) — no more memory leaks
	• FIX: NLI input format — pass (text_a, text_b) tuple, not [SEP]-concatenated string
	• FIX: Classifier max_length raised to 512 (was 256 — truncating legal clauses)
	• FIX: Risk score formula — absolute risk, not normalized by total_clauses
	• FIX: Train/inference alignment — use softmax+argmax for single-label model
	• FIX: Added missing regex fallback patterns for more CUAD categories
	• FIX: Entity extraction batching — single pipeline call instead of sequential
	• PERF: Shared model singleton via models.py module
	• PERF: LRU-bounded caches everywhere

	Carried from v4.0:
	• OCR support for scanned PDFs (docTR engine with smart native/scanned routing)
	• Contract Q&A Chatbot (RAG: embedding retrieval + HF Inference API streaming)
	• Clause Redlining (3-tier: template lookup + RAG + LLM refinement)
	• Fixed CUAD label mapping (added missing index 6)
	• Structure-aware clause splitting
	• Real NLI contradiction detection via cross-encoder model
	• ML-based Legal NER with regex fallback
	• Semantic compliance checking with negation handling
	• Improved obligation extraction with false-positive filtering
	• LLM-powered clause explanations
	• Per-session temp files (no collision)
	• Model health reporting

	Models:
	• Clause classifier: Mokshith31/legalbert-contract-clause-classification
	(LoRA adapter on nlpaueb/legal-bert-base-uncased, 41 CUAD classes)
	• Legal NER: matterstack/legal-bert-ner (token classification)
	• NLI: cross-encoder/nli-deberta-v3-base (contradiction detection)
	• Embeddings: sentence-transformers/all-MiniLM-L6-v2 (RAG retrieval)
	• OCR: docTR fast_base + crnn_vgg16_bn (scanned PDF extraction)
	• LLM: Qwen/Qwen2.5-7B-Instruct via HF Inference API (chatbot + redlining)
	"""

	import os
	import re
	import json
	import csv
	import io
	import uuid
	import tempfile
	import hashlib
	import threading
	from collections import defaultdict, OrderedDict
	from datetime import datetime
	from functools import lru_cache

	import gradio as gr
	import numpy as np

	# ── Document parsers (soft-fail) ────────────────────────────────────
	try:
	import pdfplumber
	_HAS_PDF = True
	except Exception:
	_HAS_PDF = False

	try:
	from docx import Document as DocxDocument
	_HAS_DOCX = True
	except Exception:
	_HAS_DOCX = False

	# ── PyTorch / Transformers (soft-fail) ────────────────────────────────
	_HAS_TORCH = False
	_HAS_NER_MODEL = False
	_HAS_NLI_MODEL = False

	try:
	import torch
	from transformers import (
	AutoTokenizer, AutoModelForSequenceClassification,
	AutoModelForTokenClassification, pipeline
	)
	from peft import PeftModel
	_HAS_TORCH = True
	# PERF v4.3: Limit PyTorch threads to avoid CPU thrashing under concurrent requests.
	# HF Spaces CPU-basic has 2 vCPUs. Reserve 1 thread for Gradio server.
	torch.set_num_threads(2)
	torch.set_num_interop_threads(1)
	except Exception:
	pass

	# ── ONNX Runtime (soft-fail, for quantized model) ─────────────────────
	_HAS_ORT = False
	try:
	from optimum.onnxruntime import ORTModelForSequenceClassification as _ORTModel
	_HAS_ORT = True
	except ImportError:
	pass

	# ── CrossEncoder for NLI (soft-fail) ──────────────────────────────────
	_HAS_CROSS_ENCODER = False
	try:
	from sentence_transformers import CrossEncoder as _CrossEncoder
	_HAS_CROSS_ENCODER = True
	except ImportError:
	pass

	# ── Import submodules ───────────────────────────────────────────────
	from compare import compare_contracts, render_comparison_html
	from obligations import extract_obligations, render_obligations_html
	from compliance import check_compliance, render_compliance_html
	from ocr_engine import parse_pdf_smart, get_ocr_status
	from chatbot import index_contract, chat_respond, get_chatbot_status
	from redlining import generate_redlines, render_redlines_html

	# ═══════════════════════════════════════════════════════════════════════
	# 1. CONFIGURATION — FIXED label mapping (41 labels, index 6 restored)
	# ═══════════════════════════════════════════════════════════════════════

	CUAD_LABELS = [
	"Document Name", # 0
	"Parties", # 1
	"Agreement Date", # 2
	"Effective Date", # 3
	"Expiration Date", # 4
	"Renewal Term", # 5
	"Notice Period to Terminate Renewal", # 6 ← WAS MISSING
	"Governing Law", # 7
	"Most Favored Nation", # 8
	"Non-Compete", # 9
	"Exclusivity", # 10
	"No-Solicit of Customers", # 11
	"No-Solicit of Employees", # 12
	"Non-Disparagement", # 13
	"Termination for Convenience", # 14
	"ROFR/ROFO/ROFN", # 15
	"Change of Control", # 16
	"Anti-Assignment", # 17
	"Revenue/Profit Sharing", # 18
	"Price Restriction", # 19
	"Minimum Commitment", # 20
	"Volume Restriction", # 21
	"IP Ownership Assignment", # 22
	"Joint IP Ownership", # 23
	"License Grant", # 24
	"Non-Transferable License", # 25
	"Affiliate License-Licensor", # 26
	"Affiliate License-Licensee", # 27
	"Unlimited/All-You-Can-Eat License", # 28
	"Irrevocable or Perpetual License", # 29
	"Source Code Escrow", # 30
	"Post-Termination Services", # 31
	"Audit Rights", # 32
	"Uncapped Liability", # 33
	"Cap on Liability", # 34
	"Liquidated Damages", # 35
	"Warranty Duration", # 36
	"Insurance", # 37
	"Covenant Not to Sue", # 38
	"Third Party Beneficiary", # 39
	"Other", # 40
	]

	_UNFAIR_LABELS = [
	"Limitation of liability", "Unilateral termination", "Unilateral change",
	"Content removal", "Contract by using", "Choice of law",
	"Jurisdiction", "Arbitration"
	]

	# FIX v4.2: Include regex-only labels that aren't in CUAD or Unfair lists
	_EXTRA_REGEX_LABELS = [
	"Indemnification", "Confidentiality", "Force Majeure", "Penalties"
	]

	_ALL_LABELS = CUAD_LABELS + _UNFAIR_LABELS + _EXTRA_REGEX_LABELS

	RISK_MAP = {
	# Critical
	"Uncapped Liability": "CRITICAL",
	"Arbitration": "CRITICAL",
	"IP Ownership Assignment": "CRITICAL",
	"Termination for Convenience": "CRITICAL",
	"Limitation of liability": "CRITICAL",
	"Unilateral termination": "CRITICAL",
	"Liquidated Damages": "CRITICAL",
	# High
	"Non-Compete": "HIGH",
	"Exclusivity": "HIGH",
	"Change of Control": "HIGH",
	"No-Solicit of Customers": "HIGH",
	"No-Solicit of Employees": "HIGH",
	"Unilateral change": "HIGH",
	"Content removal": "HIGH",
	"Anti-Assignment": "HIGH",
	"Notice Period to Terminate Renewal": "HIGH",
	# Medium
	"Governing Law": "MEDIUM",
	"Jurisdiction": "MEDIUM",
	"Choice of law": "MEDIUM",
	"Price Restriction": "MEDIUM",
	"Minimum Commitment": "MEDIUM",
	"Volume Restriction": "MEDIUM",
	"Non-Disparagement": "MEDIUM",
	"Most Favored Nation": "MEDIUM",
	"Revenue/Profit Sharing": "MEDIUM",
	"Warranty Duration": "MEDIUM",
	# Low
	"Document Name": "LOW",
	"Parties": "LOW",
	"Agreement Date": "LOW",
	"Effective Date": "LOW",
	"Expiration Date": "LOW",
	"Renewal Term": "LOW",
	"Joint IP Ownership": "LOW",
	"License Grant": "LOW",
	"Non-Transferable License": "LOW",
	"Affiliate License-Licensor": "LOW",
	"Affiliate License-Licensee": "LOW",
	"Unlimited/All-You-Can-Eat License": "LOW",
	"Irrevocable or Perpetual License": "LOW",
	"Source Code Escrow": "LOW",
	"Post-Termination Services": "LOW",
	"Audit Rights": "LOW",
	"Cap on Liability": "LOW",
	"Insurance": "LOW",
	"Covenant Not to Sue": "LOW",
	"Third Party Beneficiary": "LOW",
	"Other": "LOW",
	"ROFR/ROFO/ROFN": "LOW",
	"Contract by using": "LOW",
	# FIX v4.2: Added regex-only labels that were missing from RISK_MAP
	"Indemnification": "HIGH",
	"Confidentiality": "MEDIUM",
	"Force Majeure": "LOW",
	"Penalties": "HIGH",
	}

	DESC_MAP = {label: label.replace("_", " ") for label in _ALL_LABELS}
	DESC_MAP.update({
	"Limitation of liability": "Company limits or excludes liability for losses, data breaches, or service failures.",
	"Unilateral termination": "Company can terminate your account at any time without reason.",
	"Unilateral change": "Company can change terms at any time without your consent.",
	"Content removal": "Company can delete your content without notice or justification.",
	"Contract by using": "You are bound to the contract simply by using the service.",
	"Choice of law": "Governing law may differ from your country, reducing your legal protections.",
	"Jurisdiction": "Disputes must be resolved in a jurisdiction that may disadvantage you.",
	"Arbitration": "Forces disputes to arbitration instead of court. You waive your right to sue.",
	"Uncapped Liability": "No financial limit on damages the party may be liable for.",
	"Cap on Liability": "Maximum financial liability is explicitly capped.",
	"Non-Compete": "Restrictions on competing with the counter-party.",
	"Exclusivity": "Obligation to deal exclusively with one party.",
	"IP Ownership Assignment": "Intellectual property rights are transferred entirely.",
	"Termination for Convenience": "Either party may terminate without cause or notice.",
	"Governing Law": "Specifies which jurisdiction's laws apply.",
	"Non-Disparagement": "Agreement not to speak negatively about the other party.",
	"ROFR/ROFO/ROFN": "Right of First Refusal / Offer / Negotiation clause.",
	"Change of Control": "Provisions triggered by ownership or control changes.",
	"Anti-Assignment": "Restrictions on transferring contract rights to third parties.",
	"Liquidated Damages": "Pre-determined damages amount for breach of contract.",
	"Source Code Escrow": "Third-party holds source code for release under defined conditions.",
	"Post-Termination Services": "Services to be provided after the contract ends.",
	"Audit Rights": "Right to inspect records or verify compliance.",
	"Warranty Duration": "Length of time warranties remain in effect.",
	"Covenant Not to Sue": "Agreement not to bring legal action against a party.",
	"Third Party Beneficiary": "Non-party who benefits from the contract terms.",
	"Insurance": "Insurance coverage requirements.",
	"Revenue/Profit Sharing": "Revenue or profit sharing arrangements between parties.",
	"Price Restriction": "Restrictions on pricing or discounting.",
	"Minimum Commitment": "Minimum purchase or usage commitment.",
	"Volume Restriction": "Limits on volume of goods or services.",
	"License Grant": "Permission to use intellectual property.",
	"Non-Transferable License": "License that cannot be transferred to third parties.",
	"Irrevocable or Perpetual License": "License that cannot be revoked or lasts indefinitely.",
	"Unlimited/All-You-Can-Eat License": "License with no usage limits.",
	"Notice Period to Terminate Renewal": "Required notice period before automatic renewal.",
	# FIX v4.2: Added descriptions for regex-only labels
	"Indemnification": "Obligation to compensate the other party for losses or damages.",
	"Confidentiality": "Restrictions on sharing proprietary or sensitive information.",
	"Force Majeure": "Excuses performance due to extraordinary events beyond control.",
	"Penalties": "Financial penalties for breach or late performance.",
	})

	RISK_WEIGHTS = {"CRITICAL": 40, "HIGH": 20, "MEDIUM": 10, "LOW": 3}

	# FIX v4.3.1: Content-based severity refinement
	# Default RISK_MAP assigns severity by label alone. This function downgrades severity
	# when the clause text contains mitigating language (caps, carve-outs, time limits).
	_SEVERITY_MITIGATORS = {
	"IP Ownership Assignment": {
	# Downgrade from CRITICAL to HIGH if pre-existing IP is carved out
	"HIGH": re.compile(r'pre[\-\s]existing\|background\s+ip\|prior\s+(?:ip\|intellectual)', re.IGNORECASE),
	# Downgrade to MEDIUM if both carve-out AND license-back exist
	"MEDIUM": re.compile(r'(?:pre[\-\s]existing\|background\s+ip).*(?:license\|retain)', re.IGNORECASE \| re.DOTALL),
	},
	"Limitation of liability": {
	# Downgrade from CRITICAL to HIGH if there's any cap
	"HIGH": re.compile(r'shall\s+not\s+exceed\|aggregate.{0,20}(?:not\s+exceed\|limited\s+to)\|cap(?:ped)?\s+at', re.IGNORECASE),
	# Downgrade to MEDIUM if there's a reasonable cap AND exceptions for gross negligence
	"MEDIUM": re.compile(r'(?:shall\s+not\s+exceed\|limited\s+to).{0,80}(?:gross\s+negligence\|willful\|fraud)', re.IGNORECASE \| re.DOTALL),
	},
	"Termination for Convenience": {
	# Downgrade from CRITICAL to HIGH if there's a notice period
	"HIGH": re.compile(r'(?:\d+)\s+(?:day\|month\|week)s?.{0,20}(?:prior\|advance\|written)\s+notice', re.IGNORECASE),
	# Downgrade to MEDIUM if mutual termination right
	"MEDIUM": re.compile(r'either\s+party\s+may\s+terminat', re.IGNORECASE),
	},
	"Non-Compete": {
	# Downgrade from HIGH to MEDIUM if time-limited
	"MEDIUM": re.compile(r'(?:period\s+of\|for)\s+(?:\d+\|one\|two\|three\|six\|twelve)\s+(?:$\d+$\s+)?(?:month\|year)', re.IGNORECASE),
	},
	"Arbitration": {
	# Downgrade from CRITICAL to HIGH if opt-out is available
	"HIGH": re.compile(r'opt[\-\s]?out\|may\s+elect\|small\s+claims', re.IGNORECASE),
	},
	}


	def _refine_severity(label, text, default_risk):
	"""FIX v4.3.1: Refine severity based on clause content, not just label."""
	mitigators = _SEVERITY_MITIGATORS.get(label)
	if not mitigators:
	return default_risk

	# Check from lowest severity up — return the lowest matching level
	for level in ["MEDIUM", "HIGH"]:
	pattern = mitigators.get(level)
	if pattern and pattern.search(text):
	# Only downgrade, never upgrade
	level_order = {"CRITICAL": 4, "HIGH": 3, "MEDIUM": 2, "LOW": 1}
	if level_order.get(level, 0) < level_order.get(default_risk, 0):
	return level

	return default_risk

	RISK_STYLES = {
	"CRITICAL": ("#dc2626", "#fef2f2", "⚠️"),
	"HIGH": ("#ea580c", "#fff7ed", "⚡"),
	"MEDIUM": ("#ca8a04", "#fefce8", "📋"),
	"LOW": ("#16a34a", "#f0fdf4", "✓"),
	}

	# ═══════════════════════════════════════════════════════════════════════
	# FIX v4.1: Per-class thresholds aligned with single-label softmax
	# The model was trained with cross-entropy (single-label), so inference
	# now uses softmax+argmax, not sigmoid. Thresholds apply to softmax probs.
	# ═══════════════════════════════════════════════════════════════════════
	_CUAD_THRESHOLDS = {}
	_WEAK_CLASSES = {0, 1, 2, 7, 9, 21, 22, 27, 37, 38}
	for _i in range(41):
	if _i in _WEAK_CLASSES:
	_CUAD_THRESHOLDS[_i] = 0.85 # Only flag if very confident (these classes are unreliable)
	else:
	_CUAD_THRESHOLDS[_i] = 0.40 # Reasonable threshold for softmax outputs

	# ═══════════════════════════════════════════════════════════════════════
	# FIX v4.1: Bounded LRU Cache utility (replaces unbounded dicts)
	# ═══════════════════════════════════════════════════════════════════════

	class BoundedCache:
	"""Thread-safe bounded LRU cache using OrderedDict + RLock.
	FIX v4.2: Added threading.RLock to prevent race conditions under
	Gradio's concurrent request handling. OrderedDict compound operations
	(contains + setitem + move_to_end + popitem) are NOT atomic even with GIL."""
	def __init__(self, maxsize=1000):
	self._cache = OrderedDict()
	self._maxsize = maxsize
	self._lock = threading.RLock()

	def get(self, key, default=None):
	with self._lock:
	if key in self._cache:
	self._cache.move_to_end(key)
	return self._cache[key]
	return default

	def put(self, key, value):
	with self._lock:
	if key in self._cache:
	self._cache.move_to_end(key)
	self._cache[key] = value
	else:
	if len(self._cache) >= self._maxsize:
	self._cache.popitem(last=False)
	self._cache[key] = value

	def __contains__(self, key):
	with self._lock:
	return key in self._cache

	def __len__(self):
	with self._lock:
	return len(self._cache)


	# ═══════════════════════════════════════════════════════════════════════
	# 2. MODEL LOADING
	# ═══════════════════════════════════════════════════════════════════════

	cuad_tokenizer = None
	cuad_model = None
	ner_pipeline = None
	nli_model = None # FIX v4.2: CrossEncoder instead of pipeline
	_model_status = {"cuad": "not_loaded", "ner": "not_loaded", "nli": "not_loaded"}

	def _load_cuad_model():
	global cuad_tokenizer, cuad_model, _model_status
	# PERF v4.3: Try ONNX quantized model first (2-4x faster on CPU)
	onnx_model_path = os.environ.get("ONNX_MODEL_PATH", "")
	onnx_hub_id = os.environ.get("ONNX_HUB_MODEL_ID", "gaurv007/clauseguard-onnx-int8")

	if _HAS_ORT:
	for source in [onnx_model_path, onnx_hub_id]:
	if not source:
	continue
	try:
	print(f"[ClauseGuard] Trying ONNX model: {source}")
	cuad_model = _ORTModel.from_pretrained(source, file_name="model_quantized.onnx")
	cuad_tokenizer = AutoTokenizer.from_pretrained(source)
	_model_status["cuad"] = "loaded (ONNX INT8)"
	print(f"[ClauseGuard] ONNX INT8 model loaded from {source}")
	return
	except Exception as e:
	print(f"[ClauseGuard] ONNX load failed from {source}: {e}")

	# Fallback to PyTorch PEFT model
	if not _HAS_TORCH:
	print("[ClauseGuard] PyTorch not available — using regex fallback")
	_model_status["cuad"] = "unavailable"
	return
	try:
	base = "nlpaueb/legal-bert-base-uncased"
	adapter = "Mokshith31/legalbert-contract-clause-classification"
	print(f"[ClauseGuard] Loading CUAD classifier (PyTorch): {adapter}")
	cuad_tokenizer = AutoTokenizer.from_pretrained(base)
	base_model = AutoModelForSequenceClassification.from_pretrained(
	base, num_labels=41, ignore_mismatched_sizes=True
	)
	cuad_model = PeftModel.from_pretrained(base_model, adapter)
	cuad_model.eval()
	_model_status["cuad"] = "loaded (PyTorch)"
	print("[ClauseGuard] CUAD model loaded successfully (PyTorch)")
	except Exception as e:
	print(f"[ClauseGuard] CUAD model load failed: {e}")
	cuad_tokenizer = None
	cuad_model = None
	_model_status["cuad"] = f"failed: {e}"

	def _load_ner_model():
	global ner_pipeline, _model_status, _HAS_NER_MODEL
	if not _HAS_TORCH:
	_model_status["ner"] = "unavailable"
	return
	try:
	print("[ClauseGuard] Loading Legal NER model: matterstack/legal-bert-ner")
	ner_pipeline = pipeline(
	"ner",
	model="matterstack/legal-bert-ner",
	aggregation_strategy="simple",
	device=-1, # CPU
	)
	_HAS_NER_MODEL = True
	_model_status["ner"] = "loaded"
	print("[ClauseGuard] Legal NER model loaded successfully")
	except Exception as e:
	print(f"[ClauseGuard] Legal NER model load failed (using regex fallback): {e}")
	_model_status["ner"] = f"failed: {e}"

	def _load_nli_model():
	global nli_model, _model_status, _HAS_NLI_MODEL
	if not _HAS_CROSS_ENCODER:
	_model_status["nli"] = "unavailable (sentence-transformers not installed)"
	return
	try:
	print("[ClauseGuard] Loading NLI model: cross-encoder/nli-deberta-v3-base (CrossEncoder)")
	nli_model = _CrossEncoder("cross-encoder/nli-deberta-v3-base")
	_HAS_NLI_MODEL = True
	_model_status["nli"] = "loaded"
	print("[ClauseGuard] NLI CrossEncoder loaded successfully")
	except Exception as e:
	print(f"[ClauseGuard] NLI model load failed (using heuristic fallback): {e}")
	_model_status["nli"] = f"failed: {e}"

	def get_model_status_text():
	"""Return human-readable model status."""
	parts = []
	for name, status in _model_status.items():
	icon = "✅" if status == "loaded" else "⚠️" if "failed" in status else "❌"
	label = {"cuad": "Clause Classifier", "ner": "Legal NER", "nli": "NLI Contradiction"}[name]
	parts.append(f"{icon} {label}: {status}")
	return " · ".join(parts)

	# Load models at startup
	_load_cuad_model()
	_load_ner_model()
	_load_nli_model()

	# ═══════════════════════════════════════════════════════════════════════
	# 3. DOCUMENT PARSING
	# ═══════════════════════════════════════════════════════════════════════

	def parse_pdf(file_path):
	"""Smart PDF parser: native text extraction with OCR fallback for scanned PDFs."""
	text, error, method = parse_pdf_smart(file_path)
	if text:
	if method == "ocr":
	print(f"[ClauseGuard] PDF extracted via OCR ({len(text)} chars)")
	return text, None
	if error:
	return None, error
	return None, "Could not extract text from PDF. Try uploading a clearer scan or digital PDF."

	def parse_docx(file_path):
	if not _HAS_DOCX:
	return None, "DOCX parsing not available (python-docx not installed)"
	try:
	doc = DocxDocument(file_path)
	paragraphs = [p.text for p in doc.paragraphs if p.text.strip()]
	return "\n\n".join(paragraphs), None
	except Exception as e:
	return None, f"DOCX parse error: {e}"

	def parse_document(file_path):
	if file_path is None:
	return None, "No file uploaded"
	ext = os.path.splitext(file_path)[1].lower()
	if ext == ".pdf":
	return parse_pdf(file_path)
	elif ext in (".docx", ".doc"):
	return parse_docx(file_path)
	elif ext in (".txt", ".md", ".rst"):
	try:
	with open(file_path, "r", encoding="utf-8", errors="ignore") as f:
	return f.read(), None
	except Exception as e:
	return None, f"Text read error: {e}"
	else:
	return None, f"Unsupported file type: {ext}"

	# ═══════════════════════════════════════════════════════════════════════
	# 4. DETERMINISTIC CLAUSE SPLITTING
	# FIX v4.1: Bounded cache (max 500 documents) instead of unbounded dict
	# ═══════════════════════════════════════════════════════════════════════

	_chunk_cache = BoundedCache(maxsize=500)

	# FIX v4.2: Pre-compile section pattern at module level (was recompiling per call)
	_SECTION_PATTERN = re.compile(
	r'(?:^\|\n\n)'
	r'(?='
	r'\d+(?:\.\d+)*[.)]\s' # 1. 2. 3.1. 3.1)
	r'\|[A-Z]{2,}[A-Z\s]*\n' # ALL CAPS HEADERS
	r'\|$[a-z]$\s' # (a) (b) (c)
	r'\|(?:Section\|Article\|Clause)\s+\d+' # Section 1, Article 2
	r')',
	re.MULTILINE
	)

	def split_clauses(text):
	"""Deterministic, structure-aware clause splitting.
	Same input ALWAYS produces same output. Normalized text is hashed
	and cached so repeated runs on identical documents are identical."""
	normalized = re.sub(r'\s+', ' ', text.strip())
	text_hash = hashlib.sha256(normalized.encode()).hexdigest()
	cached = _chunk_cache.get(text_hash)
	if cached is not None:
	return cached

	text = re.sub(r'\n{3,}', '\n\n', text.strip())

	# First try to detect numbered sections (1., 2., 3.1, (a), etc.)
	positions = [m.start() for m in _SECTION_PATTERN.finditer(text)]

	if len(positions) >= 3:
	clauses = []
	for i, pos in enumerate(positions):
	end = positions[i + 1] if i + 1 < len(positions) else len(text)
	chunk = text[pos:end].strip()
	if len(chunk) > 30:
	if len(chunk) > 1500:
	sub_parts = chunk.split('\n\n')
	current = ""
	for sp in sub_parts:
	if len(current) + len(sp) < 1200:
	current += ("\n\n" + sp if current else sp)
	else:
	if len(current.strip()) > 30:
	clauses.append(current.strip())
	current = sp
	if len(current.strip()) > 30:
	clauses.append(current.strip())
	else:
	clauses.append(chunk)
	if positions and positions[0] > 50:
	preamble = text[:positions[0]].strip()
	if len(preamble) > 30:
	clauses.insert(0, preamble)
	result = clauses if clauses else _fallback_split(text)
	_chunk_cache.put(text_hash, result)
	return result
	else:
	result = _fallback_split(text)
	_chunk_cache.put(text_hash, result)
	return result

	def _fallback_split(text):
	"""Fallback: split on paragraph breaks and sentence boundaries."""
	paragraphs = text.split('\n\n')
	if len(paragraphs) >= 3:
	clauses = []
	for p in paragraphs:
	p = p.strip()
	if len(p) > 30:
	if len(p) > 1500:
	sents = re.split(r'(?<=[.!?])\s+(?=[A-Z])', p)
	current = ""
	for s in sents:
	if len(current) + len(s) < 1000:
	current += (" " + s if current else s)
	else:
	if len(current.strip()) > 30:
	clauses.append(current.strip())
	current = s
	if len(current.strip()) > 30:
	clauses.append(current.strip())
	else:
	clauses.append(p)
	return clauses

	parts = re.split(r'(?<=[.!?])\s+(?=[A-Z0-9(])', text)
	return [p.strip() for p in parts if len(p.strip()) > 30]

	# ═══════════════════════════════════════════════════════════════════════
	# 5. CLAUSE DETECTION
	# FIX v4.1: Use softmax (matching training) instead of sigmoid
	# FIX v4.1: max_length raised to 512 (was 256)
	# FIX v4.1: Bounded prediction cache
	# ═══════════════════════════════════════════════════════════════════════

	_HEADING_RE = re.compile(r'^\d+(?:\.\d+)*\s+[A-Z][A-Z\s&,/]+$', re.MULTILINE)

	def _strip_heading(text):
	"""Remove leading section headings that confuse the classifier."""
	lines = text.split('\n')
	if lines and _HEADING_RE.match(lines[0].strip()):
	stripped = '\n'.join(lines[1:]).strip()
	return stripped if len(stripped) > 20 else text
	return text

	_LABEL_GUARDRAILS = {
	"Liquidated Damages": re.compile(
	r'liquidated\|pre-?determined.{0,10}damage\|agreed.{0,10}sum\|penalty clause\|stipulated.{0,10}damage',
	re.IGNORECASE
	),
	"Uncapped Liability": re.compile(
	r'uncapped\|unlimited.{0,10}liabilit\|no.{0,10}(limit\|cap).{0,10}liabilit',
	re.IGNORECASE
	),
	"ROFR/ROFO/ROFN": re.compile(
	r'right\s+of\s+first\s+(?:refusal\|offer\|negotiation)\|ROFR\|ROFO\|ROFN',
	re.IGNORECASE
	),
	"Renewal Term": re.compile(
	r'renew(?:al)?\|successive\s+term\|auto(?:matic(?:ally)?)?\s*[\-\s]?renew\|non[\-\s]?renewal',
	re.IGNORECASE
	),
	# FIX v4.3.1: Revenue/Profit Sharing fires on IP assignment "rights for value" language
	"Revenue/Profit Sharing": re.compile(
	r'revenue\s+shar\|profit\s+shar\|revenue\s+split\|percentage\s+of\s+revenue\|royalt(?:y\|ies)\|gross\s+profit',
	re.IGNORECASE
	),
	# FIX v4.3.1: Minimum Commitment fires on fee schedules — require explicit minimum language
	"Minimum Commitment": re.compile(
	r'minimum\s+(?:purchase\|order\|spend\|volume\|commitment)\|take[\-\s]or[\-\s]pay\|minimum\s+annual',
	re.IGNORECASE
	),
	# FIX v4.3.1: Non-Disparagement fires on arbitration/class-waiver language
	"Non-Disparagement": re.compile(
	r'disparag\|defam\|false\s+statement\|negative\s+statement\|social\s+media\|reputat',
	re.IGNORECASE
	),
	}

	# FIX v4.3: Exclusion patterns — even if guardrail passes, exclude if contra-indicators present
	_LABEL_EXCLUSIONS = {
	"ROFR/ROFO/ROFN": re.compile(
	r'assigns?\s+to\|irrevocab(?:ly\|le)\s+assign\|all\s+right,?\s+title,?\s+and\s+interest\|work[\-\s]for[\-\s]hire',
	re.IGNORECASE
	),
	"Renewal Term": re.compile(
	r'limitation\s+of\s+liabilit\|shall\s+not\s+be\s+liable\|indemnif\|hold\s+harmless\|defend\s+and',
	re.IGNORECASE
	),
	# FIX v4.3.1: Revenue/Profit Sharing must NOT fire on IP assignment or license grant clauses
	"Revenue/Profit Sharing": re.compile(
	r'assigns?\s+to\|irrevocab(?:ly\|le)\s+assign\|work[\-\s](?:made\s+)?for[\-\s]hire\|license\s+to\s+access\|license\s+grant\|non[\-\s]exclusive\s+license',
	re.IGNORECASE
	),
	# FIX v4.3.1: Non-Disparagement must NOT fire on arbitration/dispute sections
	"Non-Disparagement": re.compile(
	r'arbitrat\|(?<!\w)aaa(?!\w)\|(?<!\w)jams(?!\w)\|class\s+action\|collective\s+(?:proceeding\|action)\|waives?\s+any\s+right\s+to\s+participate\|binding\s+arbitration',
	re.IGNORECASE
	),
	}

	# FIX v4.3: Minimum confidence thresholds per label
	_LABEL_MIN_CONFIDENCE = {
	"ROFR/ROFO/ROFN": 0.65,
	"Renewal Term": 0.70,
	"Revenue/Profit Sharing": 0.65, # FIX v4.3.1
	"Minimum Commitment": 0.65, # FIX v4.3.1
	}

	def _apply_guardrails(label, text, confidence):
	# Check minimum confidence for specific labels
	min_conf = _LABEL_MIN_CONFIDENCE.get(label)
	if min_conf and confidence < min_conf:
	return "Other", confidence * 0.2

	# Check required keywords (must be present)
	guard = _LABEL_GUARDRAILS.get(label)
	if guard and not guard.search(text):
	return "Other", confidence * 0.3

	# Check exclusion patterns (must NOT be present)
	exclusion = _LABEL_EXCLUSIONS.get(label)
	if exclusion and exclusion.search(text):
	return "Other", confidence * 0.2
	return label, confidence

	def _text_hash(text):
	return hashlib.md5(text.encode()).hexdigest()

	# FIX v4.1: Bounded prediction cache
	_prediction_cache = BoundedCache(maxsize=2000)

	def classify_cuad(clause_text):
	if cuad_model is None or cuad_tokenizer is None:
	return _classify_regex(clause_text)

	clean_text = _strip_heading(clause_text)

	h = _text_hash(clean_text[:512])
	cached = _prediction_cache.get(h)
	if cached is not None:
	return cached

	try:
	# FIX v4.1: max_length=512 (was 256 — truncating long legal clauses)
	inputs = cuad_tokenizer(
	clean_text,
	return_tensors="pt",
	truncation=True,
	max_length=512,
	padding=True
	)
	with torch.no_grad():
	logits = cuad_model(**inputs).logits

	# FIX v4.1: Use softmax (matching single-label cross-entropy training)
	# The model was trained with F.cross_entropy, so softmax is correct.
	probs = torch.softmax(logits, dim=-1)[0]

	# Get the top prediction
	top_prob, top_idx = torch.max(probs, dim=0)
	top_idx = int(top_idx)
	top_conf = float(top_prob)

	results = []

	# Primary prediction
	threshold = _CUAD_THRESHOLDS.get(top_idx, 0.40)
	if top_conf > threshold and top_idx < len(CUAD_LABELS):
	label = CUAD_LABELS[top_idx]
	conf = top_conf
	label, conf = _apply_guardrails(label, clause_text, conf)
	if not (label == "Other" and conf < 0.3):
	risk = RISK_MAP.get(label, "LOW")
	risk = _refine_severity(label, clause_text, risk)
	results.append({
	"label": label,
	"confidence": round(conf, 3),
	"risk": risk,
	"description": DESC_MAP.get(label, label),
	"source": "ml",
	})

	# Also check 2nd-best prediction if confident enough
	if len(probs) > 1:
	sorted_probs, sorted_indices = torch.sort(probs, descending=True)
	if len(sorted_probs) > 1:
	second_idx = int(sorted_indices[1])
	second_conf = float(sorted_probs[1])
	second_threshold = _CUAD_THRESHOLDS.get(second_idx, 0.40)
	if second_conf > second_threshold and second_idx < len(CUAD_LABELS):
	label2 = CUAD_LABELS[second_idx]
	conf2 = second_conf
	label2, conf2 = _apply_guardrails(label2, clause_text, conf2)
	if not (label2 == "Other" and conf2 < 0.3):
	# Only add if different from primary
	if not results or results[0]["label"] != label2:
	risk2 = RISK_MAP.get(label2, "LOW")
	risk2 = _refine_severity(label2, clause_text, risk2)
	results.append({
	"label": label2,
	"confidence": round(conf2, 3),
	"risk": risk2,
	"description": DESC_MAP.get(label2, label2),
	"source": "ml",
	})

	results.sort(key=lambda x: x["confidence"], reverse=True)

	# If no ML results, also try regex to catch what model misses
	if not results:
	results = _classify_regex(clause_text)

	_prediction_cache.put(h, results)
	return results
	except Exception as e:
	print(f"[ClauseGuard] CUAD inference error: {e}")
	return _classify_regex(clause_text)

	# ═══════════════════════════════════════════════════════════════════════
	# 5b. BATCHED CLAUSE CLASSIFICATION
	# PERF v4.3: Single forward pass for all clauses instead of one-by-one
	# ═══════════════════════════════════════════════════════════════════════

	def classify_cuad_batch(clauses, batch_size=8):
	"""Classify a batch of clauses in a single forward pass.
	PERF v4.3: Replaces sequential classify_cuad() loop.
	On CPU, batch_size=8 balances memory vs throughput."""
	if cuad_model is None or cuad_tokenizer is None:
	# Fallback to regex for all clauses
	return [_classify_regex(c) for c in clauses]

	all_results = []
	# Check cache first, collect uncached clauses
	uncached_indices = []
	uncached_texts = []
	for i, clause in enumerate(clauses):
	clean = _strip_heading(clause)
	h = _text_hash(clean[:512])
	cached = _prediction_cache.get(h)
	if cached is not None:
	all_results.append((i, cached))
	else:
	uncached_indices.append(i)
	uncached_texts.append(clean)
	all_results.append((i, None)) # placeholder

	if not uncached_texts:
	return [r for _, r in sorted(all_results)]

	# Process uncached in batches
	for batch_start in range(0, len(uncached_texts), batch_size):
	batch_texts = uncached_texts[batch_start:batch_start + batch_size]
	batch_original = [clauses[uncached_indices[batch_start + j]] for j in range(len(batch_texts))]

	try:
	inputs = cuad_tokenizer(
	batch_texts,
	return_tensors="pt",
	truncation=True,
	max_length=512,
	padding=True,
	)
	with torch.no_grad():
	logits = cuad_model(**inputs).logits

	probs = torch.softmax(logits, dim=-1)

	for j in range(len(batch_texts)):
	clause_probs = probs[j]
	original_text = batch_original[j]
	results = []

	# Primary prediction
	top_prob, top_idx = torch.max(clause_probs, dim=0)
	top_idx_int = int(top_idx)
	top_conf = float(top_prob)

	threshold = _CUAD_THRESHOLDS.get(top_idx_int, 0.40)
	if top_conf > threshold and top_idx_int < len(CUAD_LABELS):
	label = CUAD_LABELS[top_idx_int]
	conf = top_conf
	label, conf = _apply_guardrails(label, original_text, conf)
	if not (label == "Other" and conf < 0.3):
	risk = RISK_MAP.get(label, "LOW")
	risk = _refine_severity(label, original_text, risk)
	results.append({
	"label": label,
	"confidence": round(conf, 3),
	"risk": risk,
	"description": DESC_MAP.get(label, label),
	"source": "ml",
	})

	# 2nd-best prediction
	sorted_probs, sorted_indices = torch.sort(clause_probs, descending=True)
	if len(sorted_probs) > 1:
	second_idx = int(sorted_indices[1])
	second_conf = float(sorted_probs[1])
	second_threshold = _CUAD_THRESHOLDS.get(second_idx, 0.40)
	if second_conf > second_threshold and second_idx < len(CUAD_LABELS):
	label2 = CUAD_LABELS[second_idx]
	conf2 = second_conf
	label2, conf2 = _apply_guardrails(label2, original_text, conf2)
	if not (label2 == "Other" and conf2 < 0.3):
	if not results or results[0]["label"] != label2:
	risk2 = RISK_MAP.get(label2, "LOW")
	risk2 = _refine_severity(label2, original_text, risk2)
	results.append({
	"label": label2,
	"confidence": round(conf2, 3),
	"risk": risk2,
	"description": DESC_MAP.get(label2, label2),
	"source": "ml",
	})

	results.sort(key=lambda x: x["confidence"], reverse=True)

	if not results:
	results = _classify_regex(original_text)

	# Cache the result
	h = _text_hash(batch_texts[j][:512])
	_prediction_cache.put(h, results)

	# Update placeholder in all_results
	global_idx = uncached_indices[batch_start + j]
	for k, (idx, _) in enumerate(all_results):
	if idx == global_idx:
	all_results[k] = (idx, results)
	break

	except Exception as e:
	print(f"[ClauseGuard] Batch CUAD inference error: {e}")
	# Fallback to regex for this batch
	for j in range(len(batch_texts)):
	global_idx = uncached_indices[batch_start + j]
	results = _classify_regex(batch_original[j])
	for k, (idx, _) in enumerate(all_results):
	if idx == global_idx:
	all_results[k] = (idx, results)
	break

	return [r for _, r in sorted(all_results)]

	# FIX v4.1: Extended regex patterns to cover more CUAD categories
	_REGEX_PATTERNS = {
	"Limitation of liability": [r"not liable", r"shall not be (liable\|responsible)", r"in no event.*liable", r"limitation of liability", r"without warranty", r"disclaim"],
	"Unilateral termination": [r"terminat.at any time", r"suspend.account.*without", r"we may (terminat\|suspend\|discontinu)", r"right to (terminat\|suspend)"],
	"Unilateral change": [r"sole discretion", r"reserves? the right to (modify\|change\|update\|amend)", r"at any time.*without (prior )?notice", r"we may (modify\|change\|update)"],
	"Content removal": [r"remove.content.without", r"right to remove", r"we may.*remove"],
	"Contract by using": [r"by (using\|accessing).you agree", r"continued use.constitutes? acceptance"],
	"Choice of law": [r"governed by.*laws? of", r"shall be governed", r"laws of the state of"],
	"Jurisdiction": [r"exclusive jurisdiction", r"courts? of.(california\|delaware\|new york\|ireland\|england)", r"submit to.jurisdiction"],
	"Arbitration": [r"arbitrat", r"binding arbitration", r"waive.right.court", r"class action waiver"],
	"Governing Law": [r"governed by", r"laws of", r"jurisdiction of"],
	"Termination for Convenience": [r"terminat.for convenience", r"terminat.without cause", r"terminat.*at any time"],
	"Non-Compete": [r"non-compete", r"shall not compete", r"competition restriction"],
	"Exclusivity": [r"exclusive(?:ly)?(?:\s+(?:deal\|relationship\|partner\|right))", r"exclusivity"],
	"IP Ownership Assignment": [r"assign.intellectual property", r"ownership of.ip", r"all rights.*assign", r"work.?for.?hire"],
	"Uncapped Liability": [r"unlimited liability", r"uncapped", r"no.limit.liability"],
	"Cap on Liability": [r"cap on liability", r"maximum liability", r"liability.shall not exceed", r"aggregate liability.not exceed"],
	"Indemnification": [r"indemnif", r"hold harmless", r"defend.against.claim"],
	"Confidentiality": [r"confidential(?:ity)?", r"non-disclosure", r"\bnda\b"],
	"Force Majeure": [r"force majeure", r"act of god", r"beyond.*(?:reasonable\s+)?control"],
	"Penalties": [r"penalt(?:y\|ies)", r"late fee", r"default charge", r"interest on overdue"],
	# FIX v4.1: Added missing regex patterns for more CUAD categories
	"Audit Rights": [r"audit rights?", r"right to audit", r"inspect.records?", r"examination of.records?", r"access to.*books"],
	"Warranty Duration": [r"warrant(?:y\|ies).(?:period\|duration\|term\|months?\|years?)", r"warranty.shall.*(?:remain\|last\|continue)", r"limited warranty"],
	"Insurance": [r"(?:shall\|must).maintain.insurance", r"insurance.*coverage", r"policy of insurance", r"certificate of insurance"],
	"Source Code Escrow": [r"source code escrow", r"escrow.*source code", r"escrow agent"],
	"Post-Termination Services": [r"post.?termination.(?:service\|obligation\|support)", r"(?:after\|following\|upon).termination.(?:shall\|must\|will).(?:provide\|continue)"],
	"Renewal Term": [r"renew(?:al)?.term", r"auto(?:matic(?:ally)?)?.renew", r"successive.*(?:term\|period)"],
	"Notice Period to Terminate Renewal": [r"notice.(?:to\s+)?terminat.renew", r"(?:days?\|months?).(?:prior\|advance).(?:notice\|written).*(?:terminat\|renew)", r"notice of non.?renewal"],
	"Change of Control": [r"change of control", r"change in.(?:ownership\|control)", r"merger.acquisition", r"sale of.(?:all\|substantially).assets"],
	"Anti-Assignment": [r"(?:shall\|may)\s+not\s+assign", r"anti.?assignment", r"no.assignment.without.*consent"],
	"Revenue/Profit Sharing": [r"revenue.shar", r"profit.shar", r"royalt(?:y\|ies)"],
	"Liquidated Damages": [r"liquidated.damages?", r"pre.?determined.damage", r"stipulated.*damage"],
	"Covenant Not to Sue": [r"covenant not to sue", r"(?:shall\|agree).not.(?:bring\|file\|commence).*(?:action\|claim\|suit)"],
	"Joint IP Ownership": [r"joint(?:ly)?.own(?:ed\|ership)?.(?:ip\|intellectual property)", r"co.?own(?:ed\|ership)?"],
	"License Grant": [r"(?:grant\|license).(?:non.?exclusive\|exclusive\|perpetual\|irrevocable).(?:license\|right)", r"hereby grants?.*license"],
	"Non-Transferable License": [r"non.?transferable.license", r"license.(?:shall\|may)\s+not.*(?:transfer\|assign\|sublicense)"],
	"ROFR/ROFO/ROFN": [r"right of first.*(?:refusal\|offer\|negotiation)", r"ROFR", r"ROFO", r"ROFN"],
	"No-Solicit of Customers": [r"(?:shall\|must\|agree).not.solicit.customer", r"no.?solicit.customer", r"non.?solicitation.*customer"],
	"No-Solicit of Employees": [r"(?:shall\|must\|agree).not.solicit.employee", r"no.?solicit.employee", r"non.?solicitation.*employee", r"no.?hire"],
	"Non-Disparagement": [r"non.?disparagement", r"(?:shall\|must\|agree).not.(?:disparag\|defam\|make.negative)", r"not.make.*derogatory"],
	"Most Favored Nation": [r"most favou?red.nation", r"MFN", r"most favou?red.(?:customer\|pricing\|terms)"],
	"Third Party Beneficiary": [r"third.?party.beneficiar", r"no.third.?party.*beneficiar"],
	"Minimum Commitment": [r"minimum.(?:commitment\|purchase\|order\|volume\|spend)", r"(?:shall\|must).(?:purchase\|order).*(?:at least\|minimum\|no less than)"],
	"Volume Restriction": [r"volume.(?:restriction\|limitation\|cap\|ceiling)", r"(?:shall\|may).not.exceed.(?:volume\|quantity)"],
	"Price Restriction": [r"price.(?:restriction\|limitation\|ceiling\|cap\|floor)", r"(?:shall\|may).not.(?:increase\|raise\|exceed).price"],
	}

	# FIX v4.2: Pre-compile regex patterns at module level (was recompiling per call)
	_REGEX_PATTERNS_COMPILED = {}
	for _label, _pats in _REGEX_PATTERNS.items():
	_REGEX_PATTERNS_COMPILED[_label] = [re.compile(p, re.IGNORECASE) for p in _pats]

	def _classify_regex(text):
	"""Regex fallback — returns pattern match, NOT fake confidence."""
	text_lower = text.lower()
	results = []
	seen = set()
	for label, patterns in _REGEX_PATTERNS_COMPILED.items():
	for pat in patterns:
	if pat.search(text_lower):
	if label not in seen:
	risk = RISK_MAP.get(label, "MEDIUM")
	results.append({
	"label": label,
	"confidence": None,
	"risk": risk,
	"description": DESC_MAP.get(label, label),
	"source": "pattern",
	})
	seen.add(label)
	break
	return results

	# ═══════════════════════════════════════════════════════════════════════
	# 6. LEGAL NER — ML model with regex fallback
	# FIX v4.1: Batch all chunks in single pipeline call
	# ═══════════════════════════════════════════════════════════════════════

	def extract_entities(text):
	"""Extract entities using ML model (matterstack/legal-bert-ner) with regex fallback."""
	entities = []

	if _HAS_NER_MODEL and ner_pipeline is not None:
	try:
	# FIX v4.1: Create overlapping chunks but batch them in a SINGLE pipeline call
	max_text = min(len(text), 10000)
	chunks = [text[i:i+512] for i in range(0, max_text, 450)]
	offsets = list(range(0, max_text, 450))

	# Single batched pipeline call instead of sequential
	all_ner_results = ner_pipeline(chunks, batch_size=8)

	for chunk_idx, ner_results in enumerate(all_ner_results):
	offset = offsets[chunk_idx]
	for ent in ner_results:
	if ent.get("score", 0) > 0.5:
	entities.append({
	"text": ent["word"],
	"type": _map_ner_label(ent.get("entity_group", ent.get("entity", "MISC"))),
	"start": ent["start"] + offset,
	"end": ent["end"] + offset,
	"score": round(ent["score"], 3),
	"source": "ml",
	})
	except Exception as e:
	print(f"[ClauseGuard] ML NER error, falling back to regex: {e}")
	entities = _extract_entities_regex(text)
	else:
	entities = _extract_entities_regex(text)

	# FIX v4.3: Post-process ML entities to clean up WordPiece artefacts
	cleaned_entities = []
	for e in entities:
	text_val = e.get("text", "")
	# Strip WordPiece subword tokens (## prefix)
	if "##" in text_val:
	text_val = re.sub(r'##\w*', '', text_val).strip()
	text_val = re.sub(r'\s+', ' ', text_val).strip()
	# Discard entities that are too short, start/end with hyphens, or are garbled
	if len(text_val) < 2:
	continue
	if text_val.startswith("-") or text_val.endswith("-"):
	continue
	# Discard low-confidence MISC entities (almost always tokenisation artefacts)
	if e.get("type") == "MISC" and e.get("score", 1.0) < 0.6:
	continue
	# Discard entities that are mostly punctuation/symbols
	alpha_ratio = sum(1 for c in text_val if c.isalnum()) / max(len(text_val), 1)
	if alpha_ratio < 0.4:
	continue
	e["text"] = text_val
	cleaned_entities.append(e)
	entities = cleaned_entities

	# FIX v4.3: Split concatenated MONEY/QUANTITY entities
	# e.g., "usd $ 485, 000,usd $ 72, 000" → separate entities
	_CURRENCY_SPLIT = re.compile(r'(?<=[\d,])\s*(?=(?:USD\|usd\|EUR\|GBP\|\$\|£\|€))', re.IGNORECASE)
	split_entities = []
	for e in entities:
	if e.get("type") in ("MONEY", "QUANTITY") and _CURRENCY_SPLIT.search(e["text"]):
	parts = _CURRENCY_SPLIT.split(e["text"])
	for part in parts:
	part = part.strip().strip(",").strip()
	if len(part) >= 2:
	new_ent = dict(e)
	new_ent["text"] = re.sub(r'\s+', '', part) if "$" in part or "USD" in part.upper() else part
	split_entities.append(new_ent)
	else:
	split_entities.append(e)
	entities = split_entities

	# Always supplement with regex patterns for things NER often misses
	regex_ents = _extract_entities_regex(text)
	ml_spans = set()
	for e in entities:
	for pos in range(e["start"], e["end"]):
	ml_spans.add(pos)
	for re_ent in regex_ents:
	if not any(pos in ml_spans for pos in range(re_ent["start"], re_ent["end"])):
	entities.append(re_ent)

	# Deduplicate and sort
	entities.sort(key=lambda x: (x["start"], -(x["end"] - x["start"])))
	filtered = []
	last_end = -1
	for e in entities:
	if e["start"] >= last_end:
	filtered.append(e)
	last_end = e["end"]
	return filtered

	def _map_ner_label(label):
	label = label.upper()
	mapping = {
	"PER": "PERSON", "PERSON": "PERSON",
	"ORG": "PARTY", "ORGANIZATION": "PARTY",
	"LOC": "JURISDICTION", "LOCATION": "JURISDICTION",
	"GPE": "JURISDICTION", "DATE": "DATE",
	"MONEY": "MONEY", "MISC": "MISC", "LAW": "LEGAL_REF",
	}
	return mapping.get(label, label)

	def _extract_entities_regex(text):
	"""Regex-based NER fallback."""
	entities = []
	patterns = [
	(r'\b(?:January\|February\|March\|April\|May\|June\|July\|August\|September\|October\|November\|December)\s+\d{1,2},?\s+\d{4}\b', "DATE"),
	(r'\b\d{1,2}/\d{1,2}/\d{2,4}\b', "DATE"),
	(r'\b\d{1,2}-(?:Jan\|Feb\|Mar\|Apr\|May\|Jun\|Jul\|Aug\|Sep\|Oct\|Nov\|Dec)-\d{2,4}\b', "DATE"),
	(r'\b(?:Effective\|Commencement\|Expiration\|Termination)\s+Date\b', "DATE_REF"),
	(r'\$\s?\d{1,3}(?:,\d{3})(?:\.\d{2})?(?:\s(?:million\|billion\|thousand\|M\|B\|K))?', "MONEY"),
	(r'\b\d{1,3}(?:,\d{3})(?:\.\d{2})?\s(?:USD\|EUR\|GBP\|dollars\|euros\|pounds)', "MONEY"),
	(r'\b(?:USD\|EUR\|GBP)\s\d{1,3}(?:,\d{3})(?:\.\d{2})?', "MONEY"),
	(r'\b\d+(?:\.\d+)?%', "PERCENTAGE"),
	(r'\b\d+\s*(?:year\|month\|week\|day\|business day)s?\b', "DURATION"),
	(r'\b[A-Z][A-Za-z0-9\s&,]+?(?:Inc\.?\|LLC\|Ltd\.?\|Limited\|Corp\.?\|Corporation\|PLC\|GmbH\|AG\|S\.A\.?\|B\.V\.?\|L\.P\.?\|LLP)\b', "PARTY"),
	(r'\b(?:Party A\|Party B\|Disclosing Party\|Receiving Party\|Licensor\|Licensee\|Buyer\|Seller\|Tenant\|Landlord\|Employer\|Employee\|Customer\|Vendor\|Client)\b', "PARTY_ROLE"),
	(r'\b(?:State\|Commonwealth)\s+of\s+[A-Z][a-zA-Z\s]+', "JURISDICTION"),
	(r'\b(?:California\|Delaware\|New York\|Texas\|Florida\|England\|Ireland\|Germany\|France\|Singapore\|Hong Kong\|Ontario\|British Columbia)\b', "JURISDICTION"),
	(r'"([A-Z][A-Za-z\s]{1,40})"', "DEFINED_TERM"),
	(r'$(?:the\s+)?"([A-Z][A-Za-z\s]{1,40})"$', "DEFINED_TERM"),
	]
	for pat, etype in patterns:
	for m in re.finditer(pat, text, re.IGNORECASE if etype in ("DATE", "MONEY", "DURATION", "PERCENTAGE") else 0):
	txt = m.group(1) if m.lastindex else m.group()
	entities.append({
	"text": txt,
	"type": etype,
	"start": m.start(),
	"end": m.end(),
	"source": "pattern",
	})
	return entities

	# ═══════════════════════════════════════════════════════════════════════
	# 7. NLI / CONTRADICTION DETECTION
	# FIX v4.1: Pass (text_a, text_b) as dict with proper keys for
	# cross-encoder pipeline, not [SEP]-concatenated string
	# ═══════════════════════════════════════════════════════════════════════

	def _run_nli(text_a, text_b):
	"""Run NLI using CrossEncoder with correct input format.
	FIX v4.2: Use sentence_transformers.CrossEncoder.predict() which accepts
	a list of (text_a, text_b) tuples. Returns scores for [contradiction, entailment, neutral].
	The old code used pipeline("text-classification") with dict input, which was broken."""
	try:
	# CrossEncoder.predict returns numpy array of shape (n_pairs, 3)
	# Columns: [contradiction, entailment, neutral]
	scores = nli_model.predict([(text_a[:256], text_b[:256])])
	label_mapping = ["contradiction", "entailment", "neutral"]
	top_idx = int(scores[0].argmax())
	top_score = float(scores[0][top_idx])
	return [{"label": label_mapping[top_idx], "score": top_score}]
	except Exception as e:
	print(f"[ClauseGuard] NLI inference error: {e}")
	return None


	def detect_contradictions(clause_results, raw_text=""):
	"""
	Detect contradictions using:
	1. NLI cross-encoder model (semantic contradiction detection)
	2. Structural conflict detection (mutually exclusive labels)
	3. Missing critical clause detection
	"""
	contradictions = []
	labels_found = set()
	clause_texts_by_label = defaultdict(list)

	for cr in clause_results:
	labels_found.add(cr["label"])
	clause_texts_by_label[cr["label"]].append(cr.get("text", ""))

	# ── 1. Semantic NLI (if model available) ──
	if _HAS_NLI_MODEL and nli_model is not None:
	conflict_pairs = [
	("Uncapped Liability", "Cap on Liability",
	"Liability cannot be both uncapped and capped simultaneously."),
	("IP Ownership Assignment", "Joint IP Ownership",
	"IP cannot be both fully assigned and jointly owned."),
	("Exclusivity", "Non-Transferable License",
	"Exclusivity and non-transferable license may conflict."),
	]
	for label_a, label_b, explanation in conflict_pairs:
	if label_a in labels_found and label_b in labels_found:
	texts_a = clause_texts_by_label[label_a]
	texts_b = clause_texts_by_label[label_b]
	for ta in texts_a[:2]:
	for tb in texts_b[:2]:
	# FIX v4.1: Use proper NLI input format
	nli_result = _run_nli(ta, tb)
	if nli_result is None:
	continue
	for r in (nli_result if isinstance(nli_result, list) else [nli_result]):
	if r.get("label", "").lower() == "contradiction" and r.get("score", 0) > 0.6:
	contradictions.append({
	"type": "CONTRADICTION",
	"explanation": explanation,
	"severity": "HIGH",
	"clauses": [label_a, label_b],
	"confidence": round(r["score"], 3),
	"source": "nli_model",
	})

	# Also check for internal contradictions within governing law / termination
	for label in ["Governing Law", "Termination for Convenience"]:
	texts = clause_texts_by_label.get(label, [])
	if len(texts) >= 2:
	for i in range(len(texts)):
	for j in range(i + 1, min(len(texts), i + 3)):
	nli_result = _run_nli(texts[i], texts[j])
	if nli_result is None:
	continue
	for r in (nli_result if isinstance(nli_result, list) else [nli_result]):
	if r.get("label", "").lower() == "contradiction" and r.get("score", 0) > 0.6:
	contradictions.append({
	"type": "CONTRADICTION",
	"explanation": f"Conflicting {label} provisions detected — clauses contradict each other.",
	"severity": "HIGH",
	"clauses": [label],
	"confidence": round(r["score"], 3),
	"source": "nli_model",
	})
	else:
	# ── Heuristic fallback (improved) ──
	_heuristic_pairs = [
	(["Uncapped Liability"], ["Cap on Liability"],
	"Liability cannot be both uncapped and capped simultaneously."),
	(["IP Ownership Assignment"], ["Joint IP Ownership"],
	"IP cannot be both fully assigned and jointly owned."),
	]
	for group_a, group_b, explanation in _heuristic_pairs:
	found_a = any(l in labels_found for l in group_a)
	found_b = any(l in labels_found for l in group_b)
	if found_a and found_b:
	contradictions.append({
	"type": "CONTRADICTION",
	"explanation": explanation,
	"severity": "HIGH",
	"clauses": group_a + group_b,
	"source": "heuristic",
	})

	# ── 2. Missing critical clauses ──
	_REQUIRED_CLAUSE_PATTERNS = {
	"Governing Law": re.compile(
	r'govern(?:ed\|ing).{0,15}law\|applicable.{0,10}law\|laws?\s+of\s+the\s+state',
	re.IGNORECASE
	),
	"Limitation of liability": re.compile(
	r'limitation.{0,10}liabilit\|cap.{0,10}liabilit\|liabilit.{0,10}shall\s+not\s+exceed\|in\s+no\s+event.{0,20}liable',
	re.IGNORECASE
	),
	"Arbitration": re.compile(
	r'arbitrat\|AAA\|JAMS\|binding.{0,10}dispute',
	re.IGNORECASE
	),
	"Termination": re.compile(
	r'terminat(?:e\|ion\|ed)\|cancel(?:lation)?',
	re.IGNORECASE
	),
	}
	for clause_name, pattern in _REQUIRED_CLAUSE_PATTERNS.items():
	if not pattern.search(raw_text):
	contradictions.append({
	"type": "MISSING",
	"explanation": f"No '{clause_name}' clause detected in the document.",
	"severity": "MEDIUM",
	"clauses": [clause_name],
	"source": "structural",
	})

	# Deduplicate
	seen = set()
	unique = []
	for c in contradictions:
	key = (c["type"], c["explanation"])
	if key not in seen:
	seen.add(key)
	unique.append(c)

	return unique

	# ═══════════════════════════════════════════════════════════════════════
	# 8. RISK SCORING
	# FIX v4.1: Absolute risk based on findings, not normalized by doc length
	# ═══════════════════════════════════════════════════════════════════════

	def compute_risk_score(clause_results, total_clauses):
	sev_counts = {"CRITICAL": 0, "HIGH": 0, "MEDIUM": 0, "LOW": 0}
	for cr in clause_results:
	sev = cr.get("risk", "LOW")
	sev_counts[sev] += 1
	if total_clauses == 0:
	return 0, "A", sev_counts

	# FIX v4.3: Revised risk formula — scale denominator with clause count
	# to prevent small contracts from always scoring 80+.
	# The old formula used a fixed /30 denominator which meant even 2 CRITICAL
	# flags scored 73, making almost every contract grade F.
	#
	# New approach: dynamic denominator based on total clauses analysed.
	# This means risk is relative to document complexity.
	# - 1 CRITICAL in 5 clauses = high risk
	# - 1 CRITICAL in 50 clauses = moderate risk (proportionally less of the contract)
	weighted = sum(sev_counts[s] * RISK_WEIGHTS[s] for s in sev_counts)

	# Dynamic max: what if every clause were CRITICAL?
	max_possible = total_clauses * RISK_WEIGHTS["CRITICAL"]
	if max_possible == 0:
	max_possible = 1

	# Blend: 60% absolute (diminishing returns) + 40% relative (to total clauses)
	absolute_risk = 100 * (1 - (1 / (1 + weighted / 50))) # /50 instead of /30 = softer curve
	relative_risk = min(100, (weighted / max_possible) * 100)
	risk = min(100, round(0.6 * absolute_risk + 0.4 * relative_risk))

	if risk >= 70: grade = "F"
	elif risk >= 50: grade = "D"
	elif risk >= 30: grade = "C"
	elif risk >= 15: grade = "B"
	else: grade = "A"
	return risk, grade, sev_counts

	# ═══════════════════════════════════════════════════════════════════════
	# 9. MAIN ANALYSIS PIPELINE
	# ═══════════════════════════════════════════════════════════════════════

	def analyze_contract(text):
	if not text or len(text.strip()) < 50:
	return None, "Document too short (minimum 50 characters)"
	clauses = split_clauses(text)
	if not clauses:
	return None, "No clauses detected in document"

	# PERF v4.3: Batch classification — single forward pass instead of per-clause
	batch_predictions = classify_cuad_batch(clauses, batch_size=8)

	clause_results = []
	for clause, predictions in zip(clauses, batch_predictions):
	if predictions:
	for pred in predictions:
	clause_results.append({
	"text": clause,
	"label": pred["label"],
	"confidence": pred["confidence"],
	"risk": pred["risk"],
	"description": pred["description"],
	"source": pred.get("source", "unknown"),
	})
	entities = extract_entities(text)
	contradictions = detect_contradictions(clause_results, text)
	risk, grade, sev_counts = compute_risk_score(clause_results, len(clauses))
	obligations = extract_obligations(text)
	compliance = check_compliance(text)

	flagged_clause_count = len(clause_results)
	unique_flagged_texts = len(set(cr["text"] for cr in clause_results))

	result = {
	"metadata": {
	"analysis_date": datetime.now().isoformat(),
	"total_clauses": len(clauses),
	"flagged_clauses": flagged_clause_count,
	"unique_flagged": unique_flagged_texts,
	"model": get_model_status_text(),
	"text_hash": hashlib.sha256(re.sub(r'\s+', ' ', text.strip()).encode()).hexdigest()[:16],
	},
	"risk": {
	"score": risk,
	"grade": grade,
	"breakdown": sev_counts,
	},
	"clauses": clause_results,
	"entities": entities,
	"contradictions": contradictions,
	"obligations": obligations,
	"compliance": compliance,
	"raw_text": text,
	}
	return result, None

	# ═══════════════════════════════════════════════════════════════════════
	# 10. EXPORT FUNCTIONS
	# ═══════════════════════════════════════════════════════════════════════

	def export_json(result):
	if result is None:
	return None
	return json.dumps(result, indent=2, default=str)

	def export_csv(result):
	if result is None:
	return None
	output = io.StringIO()
	writer = csv.writer(output)
	writer.writerow(["Clause Text", "Label", "Risk", "Confidence", "Description", "Source"])
	for cr in result.get("clauses", []):
	conf = cr.get("confidence")
	conf_str = f"{conf:.3f}" if conf is not None else "pattern match"
	writer.writerow([
	cr.get("text", "")[:500],
	cr.get("label", ""),
	cr.get("risk", ""),
	conf_str,
	cr.get("description", ""),
	cr.get("source", ""),
	])
	return output.getvalue()

	# ═══════════════════════════════════════════════════════════════════════
	# 11. UI RENDERING
	# ═══════════════════════════════════════════════════════════════════════

	def render_summary(result):
	if result is None:
	return ""
	risk = result["risk"]
	score = risk["score"]
	grade = risk["grade"]
	breakdown = risk["breakdown"]
	grade_color = {
	"A": "#16a34a", "B": "#65a30d", "C": "#ca8a04",
	"D": "#ea580c", "F": "#dc2626",
	}.get(grade, "#6b7280")
	crit, high, med, low = breakdown["CRITICAL"], breakdown["HIGH"], breakdown["MEDIUM"], breakdown["LOW"]
	html = f"""
	<div style="font-family:system-ui,sans-serif;padding:16px;border:1px solid #e5e7eb;border-radius:12px;background:#fff;">
	<div style="text-align:center;margin-bottom:16px;">
	<div style="font-size:48px;font-weight:700;color:{grade_color};">{score}</div>
	<div style="font-size:14px;color:#6b7280;">/100 Risk Score</div>
	<div style="display:inline-block;margin-top:8px;padding:4px 16px;border-radius:20px;background:{grade_color};color:white;font-weight:600;font-size:14px;">
	Grade {grade}
	</div>
	</div>
	<div style="display:grid;grid-template-columns:1fr 1fr;gap:8px;margin-bottom:12px;">
	<div style="padding:8px;border-radius:6px;background:#fef2f2;text-align:center;">
	<div style="font-size:20px;font-weight:700;color:#dc2626;">{crit}</div>
	<div style="font-size:11px;color:#991b1b;">Critical</div>
	</div>
	<div style="padding:8px;border-radius:6px;background:#fff7ed;text-align:center;">
	<div style="font-size:20px;font-weight:700;color:#ea580c;">{high}</div>
	<div style="font-size:11px;color:#9a3412;">High</div>
	</div>
	<div style="padding:8px;border-radius:6px;background:#fefce8;text-align:center;">
	<div style="font-size:20px;font-weight:700;color:#ca8a04;">{med}</div>
	<div style="font-size:11px;color:#854d0e;">Medium</div>
	</div>
	<div style="padding:8px;border-radius:6px;background:#f0fdf4;text-align:center;">
	<div style="font-size:20px;font-weight:700;color:#16a34a;">{low}</div>
	<div style="font-size:11px;color:#166534;">Low</div>
	</div>
	</div>
	<div style="font-size:12px;color:#6b7280;text-align:center;">
	{result['metadata']['total_clauses']} clauses analyzed · {result['metadata']['flagged_clauses']} flagged
	<br><span style="font-size:10px;">{result['metadata']['model']}</span>
	</div>
	</div>
	"""
	return html

	def render_clause_cards(result):
	if result is None:
	return ""
	clauses = result.get("clauses", [])
	if not clauses:
	return '<div style="padding:24px;text-align:center;color:#6b7280;">No clauses detected.</div>'
	grouped = defaultdict(list)
	for cr in clauses:
	grouped[cr["text"]].append(cr)
	html = '<div style="font-family:system-ui,sans-serif;">'
	for text, items in grouped.items():
	max_risk = max(items, key=lambda x: {"CRITICAL":4,"HIGH":3,"MEDIUM":2,"LOW":1}[x["risk"]])["risk"]
	border, bg, icon = RISK_STYLES[max_risk]
	tags = ""
	for item in items:
	tag_bg = RISK_STYLES[item["risk"]][1]
	tag_color = RISK_STYLES[item["risk"]][0]
	conf = item.get("confidence")
	source = item.get("source", "")
	if conf is not None:
	conf_text = f"{conf:.0%}"
	else:
	conf_text = "pattern"
	source_icon = "🤖" if source == "ml" else "📝"
	tags += f'<span style="background:{tag_bg};color:{tag_color};border:1px solid {tag_color}33;padding:2px 8px;border-radius:12px;font-size:11px;font-weight:500;margin-right:4px;">{source_icon} {item["label"]} ({conf_text})</span>'
	descs = "".join(
	f'<p style="font-size:12px;color:#6b7280;margin:4px 0 0 0;">{item["description"]}</p>'
	for item in items
	)
	preview = text[:300] + ("..." if len(text) > 300 else "")
	preview = preview.replace("<", "<").replace(">", ">")
	html += f"""
	<div style="border:1px solid #e5e7eb;border-left:4px solid {border};border-radius:8px;padding:14px;margin-bottom:10px;background:#fafafa;">
	<div style="display:flex;align-items:center;gap:6px;margin-bottom:6px;">
	<span style="font-size:16px;">{icon}</span>
	<span style="font-size:12px;font-weight:600;color:{border};text-transform:uppercase;">{max_risk}</span>
	</div>
	<p style="font-size:13px;color:#374151;line-height:1.6;margin:0 0 8px 0;">{preview}</p>
	<div style="margin-bottom:6px;">{tags}</div>
	{descs}
	</div>
	"""
	html += "</div>"
	return html

	def render_entities(result):
	if result is None:
	return ""
	entities = result.get("entities", [])
	if not entities:
	return '<div style="padding:16px;color:#6b7280;">No entities detected.</div>'
	grouped = defaultdict(list)
	for e in entities:
	grouped[e["type"]].append(e["text"])
	html = '<div style="font-family:system-ui,sans-serif;">'
	for etype, texts in grouped.items():
	unique = list(dict.fromkeys(texts))[:20]
	color = {
	"DATE": "#3b82f6", "DATE_REF": "#60a5fa",
	"MONEY": "#22c55e", "PERCENTAGE": "#10b981",
	"DURATION": "#6366f1",
	"PARTY": "#8b5cf6", "PARTY_ROLE": "#a78bfa",
	"PERSON": "#ec4899",
	"JURISDICTION": "#f59e0b",
	"DEFINED_TERM": "#ec4899",
	"LEGAL_REF": "#6b7280",
	"MISC": "#9ca3af",
	}.get(etype, "#6b7280")
	items_html = "".join(
	f'<span style="display:inline-block;background:{color}15;color:{color};border:1px solid {color}40;padding:3px 10px;border-radius:6px;font-size:12px;margin:3px;">{t}</span>'
	for t in unique
	)
	html += f"""
	<div style="margin-bottom:12px;">
	<div style="font-size:12px;font-weight:600;color:#374151;margin-bottom:6px;text-transform:uppercase;">{etype}</div>
	<div>{items_html}</div>
	</div>
	"""
	html += "</div>"
	return html

	def render_contradictions(result):
	if result is None:
	return ""
	contradictions = result.get("contradictions", [])
	if not contradictions:
	return '<div style="padding:16px;color:#16a34a;">✓ No contradictions or missing clauses detected.</div>'
	html = '<div style="font-family:system-ui,sans-serif;">'
	for c in contradictions:
	sev_color = RISK_STYLES[c["severity"]][0]
	icon = "⚠️" if c["type"] == "CONTRADICTION" else "📋"
	source = c.get("source", "")
	source_badge = ""
	if source == "nli_model":
	conf = c.get("confidence", 0)
	source_badge = f'<span style="font-size:10px;background:#eff6ff;color:#3b82f6;padding:1px 6px;border-radius:4px;margin-left:8px;">🤖 NLI {conf:.0%}</span>'
	elif source == "heuristic":
	source_badge = '<span style="font-size:10px;background:#fef3c7;color:#92400e;padding:1px 6px;border-radius:4px;margin-left:8px;">📝 Heuristic</span>'
	html += f"""
	<div style="border:1px solid #e5e7eb;border-left:4px solid {sev_color};border-radius:8px;padding:12px;margin-bottom:8px;background:#fafafa;">
	<div style="display:flex;align-items:center;gap:6px;margin-bottom:4px;">
	<span>{icon}</span>
	<span style="font-size:12px;font-weight:600;color:{sev_color};">{c["type"]}</span>
	{source_badge}
	</div>
	<p style="font-size:13px;color:#374151;margin:0;">{c["explanation"]}</p>
	</div>
	"""
	html += "</div>"
	return html

	def render_document_viewer(result):
	if result is None:
	return ""
	text = result.get("raw_text", "")
	entities = sorted(result.get("entities", []), key=lambda x: x["start"])
	html_parts = []
	last_end = 0
	entity_colors = {
	"DATE": "#3b82f6", "DATE_REF": "#60a5fa", "MONEY": "#22c55e",
	"PERCENTAGE": "#10b981", "DURATION": "#6366f1", "PARTY": "#8b5cf6",
	"PARTY_ROLE": "#a78bfa", "PERSON": "#ec4899", "JURISDICTION": "#f59e0b",
	"DEFINED_TERM": "#ec4899", "LEGAL_REF": "#6b7280", "MISC": "#9ca3af",
	}
	for e in entities:
	if e["start"] >= last_end:
	plain = text[last_end:e["start"]].replace("<", "<").replace(">", ">")
	html_parts.append(plain)
	color = entity_colors.get(e["type"], "#6b7280")
	entity_text = text[e["start"]:e["end"]].replace("<", "<").replace(">", ">")
	html_parts.append(
	f'<span style="background:{color}20;color:{color};border-bottom:2px solid {color};padding:0 2px;border-radius:2px;" '
	f'title="{e["type"]}">{entity_text}</span>'
	)
	last_end = e["end"]
	if last_end < len(text):
	html_parts.append(text[last_end:].replace("<", "<").replace(">", ">"))
	return f'<div style="font-family:ui-monospace,monospace;font-size:13px;line-height:1.8;white-space:pre-wrap;padding:16px;">{"".join(html_parts)}</div>'

	# ═══════════════════════════════════════════════════════════════════════
	# 12. COMPARISON WRAPPER
	# ═══════════════════════════════════════════════════════════════════════

	def run_comparison(text_a, text_b):
	if not text_a or len(text_a.strip()) < 50:
	return "Contract A is too short", ""
	if not text_b or len(text_b.strip()) < 50:
	return "Contract B is too short", ""
	result = compare_contracts(text_a, text_b)
	return render_comparison_html(result), json.dumps(result, indent=2)

	# ═══════════════════════════════════════════════════════════════════════
	# 13. GRADIO UI
	# ═══════════════════════════════════════════════════════════════════════

	def process_upload(file):
	if file is None:
	return "", "No file uploaded"
	text, error = parse_document(file)
	if error:
	return "", error
	return text, "Document loaded successfully"

	def run_analysis(text):
	if not text or len(text.strip()) < 50:
	err_html = '<p style="color:#dc2626;padding:16px;">Document too short (minimum 50 characters)</p>'
	return [err_html] * 8 + [None, None, "", None]
	result, error = analyze_contract(text)
	if error:
	err_html = f'<p style="color:#dc2626;padding:16px;">{error}</p>'
	return [err_html] * 8 + [None, None, error, None]

	# FIXED: per-session temp files
	session_id = uuid.uuid4().hex[:8]
	json_path = os.path.join(tempfile.gettempdir(), f"clauseguard_{session_id}.json")
	csv_path = os.path.join(tempfile.gettempdir(), f"clauseguard_{session_id}.csv")

	with open(json_path, "w") as f:
	json.dump(result, f, indent=2, default=str)
	csv_content = export_csv(result)
	with open(csv_path, "w") as f:
	f.write(csv_content)

	# Generate redline suggestions (Tier 1 template + Tier 3 LLM for critical/high)
	redlines = generate_redlines(result, use_llm=True)
	redlines_html = render_redlines_html(redlines)

	return [
	render_summary(result),
	render_clause_cards(result),
	render_entities(result),
	render_contradictions(result),
	render_document_viewer(result),
	render_obligations_html(result.get("obligations", [])),
	render_compliance_html(result.get("compliance", {})),
	redlines_html,
	json_path,
	csv_path,
	"Analysis complete",
	result, # Store analysis result for chatbot
	]

	def do_clear():
	return [""] * 8 + [None, None, "", None]

	# ── Example contracts ──
	SPOTIFY_TOS = """By using the Spotify Service, you agree to be bound by these Terms of Use.

	Spotify may, in its sole discretion, modify or update these Terms of Service at any time without prior notice. Your continued use of the Service after any such changes constitutes your acceptance of the new Terms of Service.

	In no event will Spotify be liable for any indirect, incidental, special, consequential, or punitive damages, or any loss of profits or revenues, whether incurred directly or indirectly.

	Spotify reserves the right to remove or disable access to any User Content for any reason, without prior notice.

	Spotify may terminate your account or suspend your access at any time, with or without cause, with or without notice, effective immediately.

	These Terms will be governed by and construed in accordance with the laws of the State of New York.

	Any dispute shall be finally settled by arbitration in New York County. The parties waive any right to a jury trial."""

	RENTAL_AGREEMENT = """The Landlord reserves the right to enter the premises at any time without prior notice for inspection or any other purpose deemed necessary in their sole discretion.

	The Landlord shall not be liable for any damage to the Tenant's personal property, whether caused by water leaks, fire, theft, or any other cause, including the Landlord's own negligence.

	The Landlord may terminate this lease at any time with only 7 days written notice, for any reason or no reason at all.

	Any disputes arising from this lease agreement shall be resolved exclusively in the courts of the State of California, and the Tenant waives the right to a jury trial.

	The Landlord reserves the right to modify the terms of this lease at any time. Continued occupancy constitutes acceptance of the new terms."""

	NDA_SAMPLE = """NON-DISCLOSURE AGREEMENT

	This Non-Disclosure Agreement (the "Agreement") is entered into as of January 15, 2024 (the "Effective Date") by and between Acme Technologies, Inc. ("Disclosing Party") and Beta Solutions LLC ("Receiving Party").

	1. Governing Law. This Agreement shall be governed by and construed in accordance with the laws of the State of Delaware, without regard to its conflict of law principles.

	2. Term. This Agreement shall remain in effect for a period of three (3) years from the Effective Date.

	3. Termination. Either party may terminate this Agreement for convenience upon thirty (30) days prior written notice.

	4. Intellectual Property. All Confidential Information disclosed hereunder shall remain the exclusive property of the Disclosing Party. The Receiving Party hereby assigns to the Disclosing Party all right, title, and interest in any derivative works.

	5. Limitation of Liability. IN NO EVENT SHALL EITHER PARTY BE LIABLE FOR ANY INDIRECT, INCIDENTAL, SPECIAL, CONSEQUENTIAL, OR PUNITIVE DAMAGES.

	6. Indemnification. The Receiving Party shall indemnify and hold harmless the Disclosing Party from any and all claims arising from a breach of this Agreement.

	7. Non-Compete. During the term of this Agreement and for a period of two (2) years thereafter, the Receiving Party shall not engage in any business that competes with the Disclosing Party."""

	COMPLEX_CONTRACT = """MASTER SERVICE AGREEMENT

	This Master Service Agreement ("MSA") is entered into as of March 1, 2024 (the "Effective Date") by and between CloudTech Solutions, Inc., a Delaware corporation ("Provider") and Global Retail Partners LLC, a New York limited liability company ("Customer").

	1. SERVICES. Provider shall provide cloud hosting and data processing services as described in Exhibit A. Provider shall comply with all applicable laws including GDPR and CCPA.

	2. TERM AND RENEWAL. The initial term is twelve (12) months, automatically renewing for successive one (1) year periods unless terminated in accordance with Section 7.

	3. FEES AND PAYMENT. Customer shall pay a monthly fee of $25,000 within 30 days of invoice. Late payments incur a penalty of 1.5% per month. The total contract value is $300,000.

	4. LIABILITY. Provider's aggregate liability shall not exceed $1,000,000. IN NO EVENT SHALL PROVIDER BE LIABLE FOR LOST PROFITS OR CONSEQUENTIAL DAMAGES. Customer assumes all risk of data loss.

	5. INDEMNIFICATION. Each party shall indemnify the other for third-party claims arising from breach of this Agreement. Customer shall indemnify Provider for claims arising from Customer Data.

	6. INTELLECTUAL PROPERTY. Provider retains all IP rights. Customer receives a non-transferable, non-exclusive license for the term. Upon termination, Customer shall return or destroy all Provider materials within 10 business days.

	7. TERMINATION. Either party may terminate for convenience with 90 days notice. Provider may terminate immediately for non-payment. Upon termination, Customer shall pay all outstanding fees.

	8. GOVERNING LAW. This Agreement is governed by the laws of the State of Delaware. Disputes shall be resolved by binding arbitration in Wilmington, Delaware.

	9. FORCE MAJEURE. Neither party shall be liable for delays due to acts of God, war, terrorism, or government action.

	10. AUDIT RIGHTS. Customer may audit Provider's compliance annually. Provider shall provide SOC 2 Type II reports within 30 days of request.

	11. INSURANCE. Provider shall maintain general liability insurance of at least $5,000,000 and cyber liability insurance of at least $2,000,000.

	12. CONFIDENTIALITY. Both parties agree to keep Confidential Information secure for five (5) years. This obligation survives termination.

	13. ASSIGNMENT. Neither party may assign this Agreement without prior written consent. Any attempted assignment is void.

	14. THIRD PARTY BENEFICIARY. No third party shall have rights under this Agreement except as expressly provided."""

	# ═══════════════════════════════════════════════════════════════════════
	# 14. GRADIO BLOCKS
	# ═══════════════════════════════════════════════════════════════════════

	with gr.Blocks(
	title="ClauseGuard — AI Contract Analysis",
	css="""
	.gradio-container { max-width: 1600px !important; }
	"""
	) as demo:

	# ── Shared State (for chatbot RAG) ──────────────────────────────
	analysis_state = gr.State(None) # Full analysis result dict
	chunks_state = gr.State([]) # Contract text chunks for RAG
	embeddings_state = gr.State(None) # Chunk embeddings (numpy array)

	gr.HTML("""
	<div style="display:flex;align-items:center;justify-content:space-between;padding:12px 0;border-bottom:2px solid #e5e7eb;margin-bottom:16px;">
	<div>
	<h1 style="font-size:24px;font-weight:700;margin:0;color:#1f2937;">🛡️ ClauseGuard</h1>
	<p style="font-size:13px;color:#6b7280;margin:4px 0 0 0;">AI-Powered Legal Contract Analysis · 41 Clause Categories · Risk Scoring · ML NER · NLI Contradictions · Compliance · Obligations · <strong>Q&A Chatbot</strong> · <strong>Clause Redlining</strong> · <strong>OCR</strong></p>
	</div>
	<div style="font-size:12px;color:#9ca3af;">v4.3 · Precision Legal AI</div>
	</div>
	""")

	# ── Main Tabs: Analysis vs Comparison vs Chatbot ──
	with gr.Tabs():

	# ═══════ TAB 1: Single Contract Analysis ═══════
	with gr.Tab("📄 Single Contract Analysis"):
	with gr.Row():
	with gr.Column(scale=1):
	file_input = gr.File(
	label="📁 Upload Contract (PDF/DOCX/TXT)",
	file_types=[".pdf", ".docx", ".doc", ".txt", ".md"],
	)
	load_btn = gr.Button("Load Document", variant="secondary", size="sm")
	load_status = gr.Textbox(label="Status", interactive=False, lines=1)

	with gr.Column(scale=3):
	text_input = gr.Textbox(
	label="📄 Contract Text",
	placeholder="Paste contract text here, or upload a file above...\n\n💡 Scanned PDFs are automatically processed with OCR.",
	lines=14,
	max_lines=40,
	show_copy_button=True,
	)

	with gr.Column(scale=1):
	scan_btn = gr.Button("🔍 Analyze Contract", variant="primary", size="lg")
	clear_btn = gr.Button("Clear", variant="secondary", size="sm")
	status_msg = gr.Textbox(label="Analysis Status", interactive=False, lines=1)

	# ── Examples ──
	with gr.Row():
	gr.Examples(
	examples=[[SPOTIFY_TOS], [RENTAL_AGREEMENT], [NDA_SAMPLE], [COMPLEX_CONTRACT]],
	inputs=[text_input],
	label="Example Contracts",
	)

	# ── Results ──
	with gr.Row():
	with gr.Column(scale=1):
	gr.Markdown("### 📊 Risk Summary")
	summary_html = gr.HTML()

	gr.Markdown("### 📥 Export Reports")
	json_file = gr.File(label="JSON Report")
	csv_file = gr.File(label="CSV Report")

	with gr.Column(scale=3):
	with gr.Tabs():
	with gr.Tab("📄 Document"):
	doc_html = gr.HTML(label="Document Viewer")
	with gr.Tab("⚠️ Clauses (41 Categories)"):
	clauses_html = gr.HTML(label="Detected Clauses")
	with gr.Tab("🏷️ Entities"):
	entities_html = gr.HTML(label="Named Entities")
	with gr.Tab("🔍 Contradictions"):
	nli_html = gr.HTML(label="Contradictions & Missing Clauses")
	with gr.Tab("📋 Obligations"):
	obligations_html = gr.HTML(label="Obligation Tracker")
	with gr.Tab("⚖️ Compliance"):
	compliance_html = gr.HTML(label="Compliance Checker")
	with gr.Tab("✏️ Redlining"):
	redlining_html = gr.HTML(label="Clause Redlining Suggestions")

	# ═══════ TAB 2: Contract Comparison ═══════
	with gr.Tab("🔀 Compare Contracts"):
	with gr.Row():
	with gr.Column(scale=1):
	comp_file_a = gr.File(
	label="📁 Contract A (PDF/DOCX/TXT)",
	file_types=[".pdf", ".docx", ".doc", ".txt"],
	)
	comp_load_a = gr.Button("Load A", variant="secondary", size="sm")
	comp_status_a = gr.Textbox(label="Status A", interactive=False, lines=1)

	with gr.Column(scale=3):
	comp_text_a = gr.Textbox(
	label="Contract A",
	placeholder="Paste contract A here...",
	lines=12,
	show_copy_button=True,
	)

	with gr.Column(scale=1):
	comp_file_b = gr.File(
	label="📁 Contract B (PDF/DOCX/TXT)",
	file_types=[".pdf", ".docx", ".doc", ".txt"],
	)
	comp_load_b = gr.Button("Load B", variant="secondary", size="sm")
	comp_status_b = gr.Textbox(label="Status B", interactive=False, lines=1)

	with gr.Column(scale=3):
	comp_text_b = gr.Textbox(
	label="Contract B",
	placeholder="Paste contract B here...",
	lines=12,
	show_copy_button=True,
	)

	with gr.Row():
	with gr.Column(scale=1):
	comp_btn = gr.Button("🔀 Compare Contracts", variant="primary", size="lg")
	with gr.Column(scale=5):
	comp_status = gr.Textbox(label="Comparison Status", interactive=False, lines=1)

	with gr.Row():
	with gr.Column(scale=4):
	comp_result_html = gr.HTML(label="Comparison Results")
	with gr.Column(scale=2):
	comp_json = gr.JSON(label="Raw Comparison Data")

	# ═══════ TAB 3: Contract Q&A Chatbot ═══════
	with gr.Tab("💬 Contract Q&A"):
	gr.HTML("""
	<div style="padding:12px 16px;background:linear-gradient(135deg,#eff6ff,#faf5ff);border-radius:10px;margin-bottom:12px;border:1px solid #e5e7eb;">
	<div style="display:flex;align-items:center;gap:8px;margin-bottom:6px;">
	<span style="font-size:20px;">💬</span>
	<h3 style="margin:0;font-size:16px;color:#1f2937;">Contract Q&A Chatbot</h3>
	</div>
	<p style="font-size:12px;color:#6b7280;margin:0;line-height:1.5;">
	Ask questions about your analyzed contract. The chatbot uses <strong>RAG</strong> (Retrieval-Augmented Generation)
	to find relevant clauses and generate accurate answers grounded in your contract text.
	<br>
	<strong>Step 1:</strong> Analyze a contract in the "📄 Single Contract Analysis" tab.
	<strong>Step 2:</strong> Come here and ask questions!
	</p>
	</div>
	""")

	chatbot_index_status = gr.Textbox(
	label="📡 Chatbot Index Status",
	interactive=False,
	lines=1,
	value="⏳ No contract indexed yet — analyze a contract first",
	)

	def _chatbot_fn(message, history, chunks, embeddings, analysis):
	"""Wrapper for ChatInterface fn signature."""
	yield from chat_respond(message, history, chunks, embeddings, analysis)

	gr.ChatInterface(
	fn=_chatbot_fn,
	type="messages",
	additional_inputs=[chunks_state, embeddings_state, analysis_state],
	examples=[
	["What are the main risks in this contract?"],
	["Who are the parties involved?"],
	["What happens if the contract is terminated?"],
	["Are there any liability limitations?"],
	["What are my obligations under this contract?"],
	["Is there an arbitration clause?"],
	["What is the governing law?"],
	["Summarize the key terms in plain language."],
	],
	title="",
	description="",
	)

	# ── Events ──
	def _load_file(file):
	text, err = parse_document(file) if file else ("", "No file")
	if err and not text:
	return "", err
	return text, "Loaded successfully" if not err else err

	def _analysis_and_index(text):
	"""Run analysis AND index for chatbot in one call."""
	# Run the standard analysis
	analysis_outputs = run_analysis(text)

	# Index for chatbot (uses the raw text)
	chunks, embeddings, index_status = index_contract(text)

	# analysis_outputs has 12 items: 8 HTML + json_path + csv_path + status + result
	# We need to add: chunks_state, embeddings_state, chatbot_index_status
	return analysis_outputs + [chunks, embeddings, index_status]

	load_btn.click(_load_file, inputs=[file_input], outputs=[text_input, load_status])
	comp_load_a.click(_load_file, inputs=[comp_file_a], outputs=[comp_text_a, comp_status_a])
	comp_load_b.click(_load_file, inputs=[comp_file_b], outputs=[comp_text_b, comp_status_b])

	scan_btn.click(
	_analysis_and_index,
	inputs=[text_input],
	outputs=[
	summary_html, clauses_html, entities_html, nli_html,
	doc_html, obligations_html, compliance_html, redlining_html,
	json_file, csv_file, status_msg, analysis_state,
	chunks_state, embeddings_state, chatbot_index_status,
	],
	api_name="analyze",
	)

	clear_btn.click(
	lambda: [""] * 8 + [None, None, "", None, [], None, "⏳ No contract indexed"],
	outputs=[
	summary_html, clauses_html, entities_html, nli_html,
	doc_html, obligations_html, compliance_html, redlining_html,
	json_file, csv_file, status_msg, analysis_state,
	chunks_state, embeddings_state, chatbot_index_status,
	]
	)

	comp_btn.click(
	run_comparison,
	inputs=[comp_text_a, comp_text_b],
	outputs=[comp_result_html, comp_json],
	api_name="compare",
	)

	gr.HTML("""
	<div style="margin-top:24px;padding:16px 0;border-top:1px solid #e5e7eb;text-align:center;">
	<p style="font-size:11px;color:#9ca3af;">
	⚠️ Not legal advice. For informational purposes only.
	· Classifier: <a href="https://huggingface.co/gaurv007/clauseguard-onnx-int8" style="color:#6b7280;">Legal-BERT ONNX INT8 (41 CUAD classes)</a>
	· NER: <a href="https://huggingface.co/matterstack/legal-bert-ner" style="color:#6b7280;">Legal-BERT NER</a>
	· NLI: <a href="https://huggingface.co/cross-encoder/nli-deberta-v3-base" style="color:#6b7280;">DeBERTa-v3 NLI</a>
	· LLM: <a href="https://huggingface.co/Qwen/Qwen2.5-7B-Instruct" style="color:#6b7280;">Qwen2.5-7B</a>
	· OCR: <a href="https://github.com/mindee/doctr" style="color:#6b7280;">docTR</a>
	· Embeddings: <a href="https://huggingface.co/BAAI/bge-small-en-v1.5" style="color:#6b7280;">BGE-small-en</a>
	· Dataset: <a href="https://huggingface.co/datasets/theatticusproject/cuad-qa" style="color:#6b7280;">CUAD</a>
	· <a href="https://huggingface.co/spaces/gaurv007/ClauseGuard" style="color:#6b7280;">ClauseGuard Space</a>
	</p>
	</div>
	""")

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