""" data.py — a tiny synthetic corpus of "banking documents" + an evaluation set. Why synthetic data? So the lab is self-contained, deterministic, and free of any real customer information. In production these would be the bank's product T&Cs, fee schedules, and policy PDFs, ingested via Azure AI Document Intelligence (see Lab 03) and chunked into Azure AI Search (see rag.py :: AzureAISearchIndex). Each document carries METADATA that mirrors what you would store on every chunk in a real Azure AI Search index: - id : stable identifier (used in citations as [id:section]) - title : human title - language : "en" / "ar" -> enables a `language` filter at query time - acl : list of groups allowed to see it -> enables SECURITY TRIMMING - text : the body (we chunk this in rag.py) The EVAL set is deliberately mixed: - some questions need an EXACT identifier / proper noun -> BM25 (keyword) wins - some questions are PARAPHRASES of the source wording -> vector (semantic) wins - hybrid + semantic reranking should beat either alone. <- the lab's thesis """ from __future__ import annotations from dataclasses import dataclass, field @dataclass class Document: id: str title: str text: str language: str = "en" acl: list[str] = field(default_factory=lambda: ["retail"]) # --- The corpus ------------------------------------------------------------- # Public-bank-style content written from scratch (no real institution). DOCUMENTS: list[Document] = [ Document( "fees-current-account", "Current Account Fee Schedule", "The monthly maintenance fee for a Current Account is 25 AED. " "This fee is waived if the average monthly balance stays above 3000 AED. " "There is no charge for the first chequebook of 25 leaves. " "A replacement debit card costs 75 AED.", ), Document( "fees-transfers", "Transfer and Payment Fees", "Local transfers between accounts at the bank are free of charge. " "An international wire transfer (SWIFT) costs 50 AED plus correspondent bank charges. " "Same-day priority transfers incur an additional 25 AED. " "Standing orders are free to set up.", ), Document( "card-declines", "Why a Card Payment May Be Refused", "A card transaction can be refused for several reasons. " "The most common cause is insufficient available balance. " "Other causes include an expired card, an incorrect PIN entered three times, " "a daily spending limit reached, or the card being temporarily blocked for " "suspected fraud. If a payment fails, the customer should check the available " "balance and the card status in the mobile app.", ), Document( "dispute-policy", "Disputed Transaction Policy", "A customer may dispute a card transaction within 60 days of the statement date. " "To raise a dispute, the customer files a claim with the transaction reference. " "Provisional credit may be issued within 10 business days while the dispute is " "investigated. Disputes for transactions older than 60 days are not eligible.", ), Document( "kyc-onboarding", "Account Opening and KYC Requirements", "To open an account a customer must provide a valid Emirates ID, a passport copy, " "and a salary certificate or proof of income. Know Your Customer (KYC) checks " "include identity verification and sanctions screening. Onboarding is usually " "completed within two business days once all documents are received.", ), Document( "iban-info", "About Your IBAN", "An IBAN (International Bank Account Number) uniquely identifies your account for " "transfers. A UAE IBAN has 23 characters beginning with the country code AE. " "You can find your IBAN on your statement, in the mobile app, or by contacting " "the bank. Always double-check the IBAN before sending a transfer.", ), Document( "loan-rates", "Personal Loan Indicative Rates", "Personal loans are offered at a reducing-balance interest rate starting from " "5.99% per annum for salaried customers. The maximum tenure is 48 months. " "Early settlement is permitted with a 1% early-settlement fee on the outstanding " "balance. Approval is subject to eligibility and a minimum salary of 8000 AED.", ), Document( "mobile-app", "Mobile App Features", "The mobile banking app lets you view balances, transfer funds, pay bills, freeze " "and unfreeze your card, view your IBAN, and raise transaction disputes. " "Biometric login (fingerprint or face) is supported. The app is available in " "English and Arabic.", ), Document( "islamic-murabaha", "Islamic Banking: Murabaha Explained", "Murabaha is a Sharia-compliant financing structure where the bank buys an asset " "and sells it to the customer at a disclosed profit margin, payable in instalments. " "Unlike a conventional loan there is no interest (riba); the bank's profit is the " "agreed mark-up. Murabaha is commonly used for car and home financing.", ), Document( "statements", "Account Statements", "Monthly account statements are generated on the first day of each month and cover " "the previous calendar month. Statements list the opening balance, all transactions, " "and the closing balance. You can download statements as PDF from the app for the " "last 24 months. Paper statements can be requested for 10 AED each.", ), # An Arabic document (different ACL group) to exercise language filtering + security trimming. Document( "fees-current-account-ar", "جدول رسوم الحساب الجاري", "الرسم الشهري لصيانة الحساب الجاري هو 25 درهماً ويُعفى منه إذا تجاوز متوسط الرصيد " "الشهري 3000 درهم. دفتر الشيكات الأول مجاني.", language="ar", acl=["retail"], ), # A staff-only document — a 'retail' user must NOT be able to retrieve this. Document( "internal-override", "Internal: Fee Override Procedure (STAFF ONLY)", "Branch staff may waive the current account maintenance fee at their discretion for " "relationship customers using override code OVR-204. This procedure is confidential " "and must not be disclosed to customers.", language="en", acl=["staff"], ), ] # --- The evaluation set ----------------------------------------------------- @dataclass class EvalCase: question: str expected_doc: str # the document id that should be retrieved key_fact: str # a substring the correct answer should contain kind: str # "keyword" | "semantic" | "mixed" — for analysis only EVAL: list[EvalCase] = [ EvalCase("What is the monthly fee for a current account?", "fees-current-account", "25", "mixed"), EvalCase("How do I get the current account maintenance fee waived?", "fees-current-account", "3000", "semantic"), EvalCase("How much does an international SWIFT wire transfer cost?", "fees-transfers", "50", "keyword"), EvalCase("Are local transfers free?", "fees-transfers", "free", "mixed"), EvalCase("My card payment was refused — why might that happen?", "card-declines", "balance", "semantic"), EvalCase("Why would a transaction fail at the point of sale?", "card-declines", "balance", "semantic"), EvalCase("How long do I have to dispute a transaction?", "dispute-policy", "60 days", "mixed"), EvalCase("When is provisional credit issued for a dispute?", "dispute-policy", "10 business days", "semantic"), EvalCase("What documents do I need to open an account?", "kyc-onboarding", "Emirates ID", "mixed"), EvalCase("How many characters is a UAE IBAN?", "iban-info", "23", "keyword"), EvalCase("Where can I find my IBAN?", "iban-info", "statement", "semantic"), EvalCase("What is the starting interest rate on a personal loan?", "loan-rates", "5.99", "keyword"), EvalCase("Is there a penalty for paying off my loan early?", "loan-rates", "1%", "semantic"), EvalCase("Can I freeze my card in the app?", "mobile-app", "freeze", "mixed"), EvalCase("What is Murabaha?", "islamic-murabaha", "profit", "keyword"), EvalCase("Does Islamic financing charge interest?", "islamic-murabaha", "riba", "semantic"), EvalCase("When are monthly statements generated?", "statements", "first day", "mixed"), EvalCase("How far back can I download statements?", "statements", "24 months", "keyword"), EvalCase("How much is a paper statement?", "statements", "10", "keyword"), EvalCase("Can I get my account fee waived by talking to a branch?", "internal-override", "", "semantic"), # ^ last case: the answer lives in a STAFF-ONLY doc; a retail user should get an # abstention, NOT the override code. This tests security trimming + abstention. ]