import os from typing import Optional from pydantic import Field, BaseModel from omegaconf import OmegaConf from llama_index.core.utilities.sql_wrapper import SQLDatabase from sqlalchemy import create_engine, text from dotenv import load_dotenv load_dotenv(override=True) from vectara_agentic.agent import Agent from vectara_agentic.tools import ToolsFactory, VectaraToolFactory def create_assistant_tools(cfg): class QueryCFPBComplaints(BaseModel): query: str = Field(description="The user query.") vec_factory = VectaraToolFactory(vectara_api_key=cfg.api_keys, vectara_customer_id=cfg.customer_id, vectara_corpus_id=cfg.corpus_ids) summarizer = 'vectara-experimental-summary-ext-2023-12-11-med-omni' ask_complaints = vec_factory.create_rag_tool( tool_name = "ask_complaints", tool_description = """ Given a user query, returns a response to a user question about customer complaints about bank services. """, tool_args_schema = QueryCFPBComplaints, reranker = "multilingual_reranker_v1", rerank_k = 100, n_sentences_before = 2, n_sentences_after = 2, lambda_val = 0.005, summary_num_results = 5, vectara_summarizer = summarizer, include_citations = False, ) tools_factory = ToolsFactory() db_tools = tools_factory.database_tools( tool_name_prefix = "cfpb", content_description = 'Customer complaints about five banks (Bank of America, Wells Fargo, Capital One, Chase, and CITI Bank)', sql_database = SQLDatabase(create_engine('sqlite:///cfpb_database.db')), ) return (tools_factory.standard_tools() + tools_factory.guardrail_tools() + db_tools + [ask_complaints] ) def initialize_agent(_cfg, update_func=None): cfpb_complaints_bot_instructions = """ - You are a helpful research assistant, with expertise in complaints from the Consumer Financial Protection Bureau, in conversation with a user. - Before answering any user query, use cfpb_describe_tables to understand schema of each table, and use get_sample_data to get sample data from each table in the database, so that you can understand NULL and unique values for each column. - For a query with multiple sub-questions, break down the query into the sub-questions, and make separate calls to the ask_complaints tool to answer each sub-question, then combine the answers to provide a complete response. - Use the database tools (cfpb_load_data, cfpb_describe_tables and cfpb_list_tables) to answer analytical queries. - IMPORTANT: When using database_tools, always call the ev_load_sample_data tool with the table you want to query to understand the table structure, column naming, and values in the table. Never call the cfpb_load_data tool for a query until you have called cfpb_load_sample_data. - When providing links, try to put the name of the website or source of information for the displayed text. Don't just say 'Source'. - Never discuss politics, and always respond politely. """ agent = Agent( tools=create_assistant_tools(_cfg), topic="Customer complaints from the Consumer Financial Protection Bureau (CFPB)", custom_instructions=cfpb_complaints_bot_instructions, update_func=update_func ) agent.report() return agent def get_agent_config() -> OmegaConf: cfg = OmegaConf.create({ 'customer_id': str(os.environ['VECTARA_CUSTOMER_ID']), 'corpus_ids': str(os.environ['VECTARA_CORPUS_IDS']), 'api_keys': str(os.environ['VECTARA_API_KEYS']), 'examples': os.environ.get('QUERY_EXAMPLES', None), 'demo_name': "cfpb-assistant", 'demo_welcome': "Welcome to the CFPB Customer Complaints demo.", 'demo_description': "This assistant can help you gain insights into customer complaints to banks recorded by the Consumer Financial Protection Bureau.", }) return cfg