[ { "id": 1, "query": "What textile products does TEXBase deal with?", "expected_answer": "Yarn, cotton, polyester, and fabric sourcing.", "category": "general_knowledge", "function_tested": "none" }, { "id": 2, "query": "What is the current Brent Oil price trend?", "expected_answer": "Retrieves price from BrentOIL/brent_oil.json.", "category": "rag_retrieval", "function_tested": "Scrapping_BrentOil.main" }, { "id": 3, "query": "Generate a cold email for a company called Nike", "expected_answer": "Generates a personalized outreach email.", "category": "tool_usage", "function_tested": "ReviewAndHeaderAgent.generate_cold_email" }, { "id": 4, "query": "What is the cotton price in Pakistan today?", "expected_answer": "Retrieves price from CottonPakistan/cotton_pk.json.", "category": "rag_retrieval", "function_tested": "cotton_pk_scrapper.main" }, { "id": 5, "query": "Process a purchase order image and extract line items", "expected_answer": "Extracts structured PO data from image.", "category": "tool_usage", "function_tested": "po_processor.process_po" }, { "id": 6, "query": "Show me all emails under review", "expected_answer": "Queries emails_under_review database.", "category": "database_query", "function_tested": "db_manager.get_emails" }, { "id": 7, "query": "What is the current forex rate for USD to PKR?", "expected_answer": "Retrieves rate from Forex/forex_data.json.", "category": "rag_retrieval", "function_tested": "Scrapping.main" }, { "id": 8, "query": "Send the approved email with ID 5", "expected_answer": "Triggers Gmail API send.", "category": "tool_usage", "function_tested": "Email_sender.send_email_from_database" }, { "id": 9, "query": "Calculate risk factors for the textile market", "expected_answer": "Aggregates indicators into risk_factors.json.", "category": "tool_usage", "function_tested": "Stats.compute_risk" }, { "id": 10, "query": "Research ZANEROBE contact info", "expected_answer": "Uses Gemini Search to find contacts.", "category": "tool_usage", "function_tested": "brand_data_extraction.research" }, { "id": 11, "query": "Generate follow-up for no-response leads", "expected_answer": "Creates contextual follow-up email.", "category": "tool_usage", "function_tested": "Handling_FollowUp.generate_followup" }, { "id": 12, "query": "Categorize incoming email from supplier", "expected_answer": "Classifies email as inquiry/order/other.", "category": "tool_usage", "function_tested": "read_inbox.categorize_email" }, { "id": 13, "query": "Map Excel columns: 'Lead Name' to 'Company Name'", "expected_answer": "Correct mapping decided by LLM.", "category": "unit_test", "function_tested": "Excel_Processor.get_column_mapping" }, { "id": 14, "query": "Initialize outreach database", "expected_answer": "Creates tables for outreach_companies.", "category": "unit_test", "function_tested": "Excel_Processor.init_db" }, { "id": 15, "query": "Format email body with HTML paragraphs", "expected_answer": "Correct insertion of

and
tags.", "category": "unit_test", "function_tested": "Email_sender.format_html_body" }, { "id": 16, "query": "Update client email in database", "expected_answer": "Records updated for specific company_id.", "category": "database_query", "function_tested": "update_client_email.update" }, { "id": 17, "query": "Validate email address 'test@texbase.com'", "expected_answer": "Returns valid status.", "category": "unit_test", "function_tested": "Email_Validator.validate" }, { "id": 18, "query": "Fetch latest 5 messages from Inbox", "expected_answer": "Retrieves messages via Gmail API.", "category": "tool_usage", "function_tested": "read_inbox.fetch_messages" }, { "id": 19, "query": "Extract vendor name from PO scan", "expected_answer": "Gemini identifies vendor from OCR text.", "category": "unit_test", "function_tested": "PO_Processor.parse_po_details" }, { "id": 20, "query": "Check if user 'Asad' exists in system", "expected_answer": "Boolean return from users database.", "category": "database_query", "function_tested": "personeldata.check_user_exists" }, { "id": 21, "query": "Analyze weather impact on Sindh cotton", "expected_answer": "Correlates rain forecast with harvest risk.", "category": "rag_retrieval", "function_tested": "Weather_LLM_Strategy.analyze" }, { "id": 22, "query": "Scrape China yarn index from web", "expected_answer": "Playwright extracts latest index.", "category": "rag_retrieval", "function_tested": "Yarn_index_china.main" }, { "id": 23, "query": "Update pipeline status to 'Processing'", "expected_answer": "Status flag set in SQLite.", "category": "database_query", "function_tested": "OutreachAgent.update_status" }, { "id": 24, "query": "Draft reply to pricing inquiry", "expected_answer": "Generates professional response based on Stats.", "category": "tool_usage", "function_tested": "draft_reply.compose" }, { "id": 25, "query": "Move email from Review to Sent", "expected_answer": "Atomically updates database records.", "category": "database_query", "function_tested": "Send_email_db.approve_and_move" }, { "id": 26, "query": "Extract total amount from PO", "expected_answer": "Regex/LLM extracts currency value.", "category": "unit_test", "function_tested": "PO_Processor.extract_total" }, { "id": 27, "query": "Search knowledge base for 'Reactive Dyes'", "expected_answer": "Retrieves relevant docs from ChromaDB.", "category": "rag_retrieval", "function_tested": "vector_store.search" }, { "id": 28, "query": "Generate PDF summary of research", "expected_answer": "Writes FPDF document to disk.", "category": "unit_test", "function_tested": "researchBrain.generate_pdf" }, { "id": 29, "query": "Identify HS Code 5201 from description", "expected_answer": "Correct code for raw cotton.", "category": "unit_test", "function_tested": "Information_parser.extract_hs" }, { "id": 30, "query": "Verify JWT token validity", "expected_answer": "Token authenticated against secret.", "category": "security", "function_tested": "server.verify_token" }, { "id": 31, "query": "Handle SQL injection in lead import", "expected_answer": "Sanitizes input, prevents execution.", "category": "security", "function_tested": "Excel_Processor.save_records" }, { "id": 32, "query": "Log system event to Audit table", "expected_answer": "New entry created with timestamp.", "category": "database_query", "function_tested": "server.log_audit" }, { "id": 33, "query": "Refresh Google OAuth tokens", "expected_answer": "Obtains new access_token from refresh_token.", "category": "tool_usage", "function_tested": "Email_sender.refresh_token" }, { "id": 34, "query": "Categorize PDF as 'Invoice' or 'Quotation'", "expected_answer": "Classification based on keywords.", "category": "unit_test", "function_tested": "PO_Processor.classify_doc" }, { "id": 35, "query": "Extract shipment history from ImportYeti", "expected_answer": "Playwright scrapes brand profile tables.", "category": "rag_retrieval", "function_tested": "scrape_importyeti.parse" }, { "id": 36, "query": "Calculate KIBOR carry cost", "expected_answer": "Returns percentage based on 6M rate.", "category": "unit_test", "function_tested": "Stats.calculate_kibor" }, { "id": 37, "query": "Parse JSON string from LLM output", "expected_answer": "Removes markdown fences and validates.", "category": "unit_test", "function_tested": "ReviewAndHeaderAgent.clean_json" }, { "id": 38, "query": "Update follow-up count in database", "expected_answer": "Increments follow_up_num by 1.", "category": "database_query", "function_tested": "approve_followup_db.update" }, { "id": 39, "query": "Detect duplicate company in Excel", "expected_answer": "Skips existing website/name pairs.", "category": "unit_test", "function_tested": "Excel_Processor.check_duplicates" }, { "id": 40, "query": "Compose subject line for 'Denim' pitch", "expected_answer": "Optimized subject for open rates.", "category": "unit_test", "function_tested": "EmailGenerator.get_subject" } ]