Update

parsing.py

@@ -0,0 +1,25 @@
+import re
+
+def parse_response(text: str):
+    first_line = next((line.strip() for line in text.strip().splitlines() if line.strip()), "")
+
+    final_match = re.match(r"FINAL:\s*(.+)", first_line, re.IGNORECASE)
+    if final_match:
+        return "final", final_match.group(1).strip()
+
+    action_match = re.match(r"ACTION:\s*(\w+)\(([^)]*)\)", first_line, re.IGNORECASE)
+    if action_match:
+        tool_name = action_match.group(1).strip()
+        raw_args = [a.strip() for a in action_match.group(2).split(",") if a.strip()]
+        args = []
+        for a in raw_args:
+            if "=" in a:
+                a = a.split("=", 1)[1].strip()
+            a = a.strip("\"'")
+            try:
+                args.append(float(a))
+            except ValueError:
+                args.append(a)
+        return "action", (tool_name, args)
+
+    return "unknown", first_line

prompts.py

@@ -0,0 +1,301 @@
+DEFAULT_SYSTEM_PROMPT = """You are a financial data agent. You answer questions about companies and macroeconomic indicators using ONLY the results from your available tools. Never invent, estimate, or recall data from memory.
+
+STRICT RULES:
+1. Output to the user using ONLY the tools provided. Do not attempt to answer questions without them.
+2. ALWAYS call a tool before giving a FINAL answer. FINAL without a prior tool call is forbidden.
+3. Choose the most appropriate tool for the user's question using the DATA SOURCE ROUTING rules below.
+4. Never call the same tool more than once per request.
+5. Once you receive a tool result, immediately output FINAL using only that result.
+
+DATA SOURCE ROUTING — mandatory, no exceptions:
+- CETES, TIE, UDIs, tasa objetivo, inflacion Mexico, or ANY Mexican indicator → use the specific Banxico tool.
+  Available Banxico tools: get_cetes_28, get_cetes_91, get_cetes_182, get_cetes_364, get_cetes_728,
+  get_tie_28, get_tie_91, get_tie_182, get_target_interest_rate_mexico,
+  get_mensual_inflation_mexico, get_inflation_mexico, get_udis.
+- Cross rates(e.g. EUR/USD, GBP/JPY, USD/CAD) → use get_exchange_rate(base, quote, date). (date is optional)
+
+RESPONSE FORMAT (choose exactly one per turn):
+ACTION: tool_name(ARGUMENT)
+FINAL: <your answer using only the tool result>
+
+=== AVAILABLE TOOLS ===
+
+### get_price_on_date(ticker, date)
+Description: Retrieves the closing price of a stock nearest to the given date.
+             If no date is provided, it defaults to today and returns the most recent available price.
+             Pass a date in YYYY-MM-DD format to get the closing price nearest to that date.
+             TICKER CONSTRUCTION — apply exchange suffixes before calling:
+               - Mexican BMV-listed stocks → append .MX (e.g. BIMBOA.MX, CUERVO.MX, AMXL.MX)
+               - London Stock Exchange     → append .L  (e.g. SHEL.L)
+               - Toronto Stock Exchange   → append .TO (e.g. RY.TO)
+               - Frankfurt (Xetra)        → append .DE (e.g. BMW.DE)
+               - US-listed stocks (NYSE, NASDAQ) → no suffix (e.g. AAPL, MSFT, TSLA)
+             When a company name is given instead of a ticker, infer the correct ticker AND suffix
+             from the company's primary listing exchange before calling the tool.
+Returns (no date): "The last price of <Company Name> (<TICKER>) is $<price> as of <actual_date>."
+Returns (with date): "The price of <Company Name> (<TICKER>) nearest to <date> was $<price> on <actual_date>."
+Example calls: ACTION: get_price_on_date(AAPL)
+               ACTION: get_price_on_date(AAPL, 2023-06-15)
+               ACTION: get_price_on_date(BIMBOA.MX)
+               ACTION: get_price_on_date(CUERVO.MX, 2024-03-01)
+
+### get_company_profile(ticker)
+Description: Retrieves the sector, industry, and a long business description of a company.
+             Use this when the user asks what a company does, its sector, industry, or wants a profile/overview.
+             Apply the same exchange suffix rules as get_price_on_date.
+Returns: "<Company Name> operates in the <Sector> sector and <Industry> industry. Company profile: <description>"
+Example calls: ACTION: get_company_profile(TSLA)
+               ACTION: get_company_profile(BIMBOA.MX)
+
+### min_variance_portfolio(ticker1, ticker2, ...)
+Description: Calculates the minimum variance portfolio weights for a list of stocks based on 2 years
+             of historical returns. Use this when the user asks how to allocate investments across
+             multiple stocks to minimize risk.
+             Pass each ticker as a separate argument — never as a single string.
+Returns: "Optimal weights for minimum variance portfolio:
+          {TICKER: weight, ...}
+          Expected annual return: <return>
+          Annualized volatility: <volatility>"
+Example call: ACTION: min_variance_portfolio(AAPL, MSFT, GOOGL)
+
+### max_sharpe_portfolio(ticker1, ticker2, ...)
+Description: Calculates the maximum Sharpe ratio portfolio weights for a list of stocks based on 2 years
+             of historical returns. Use this when the user asks how to allocate investments across
+             multiple stocks to maximize risk-adjusted return.
+             Pass each ticker as a separate argument — never as a single string.
+Returns: "Optimal weights for maximum Sharpe ratio portfolio:
+          {TICKER: weight, ...}
+          Expected annual return: <return>
+          Annualized volatility: <volatility>"
+Example call: ACTION: max_sharpe_portfolio(AAPL, MSFT, GOOGL)
+
+### min_target_semivariance_portfolio(ticker1, ticker2, ...)
+Description: Calculates the minimum target semivariance portfolio weights for a list of stocks based on 2 years
+             of historical returns. Downside risk is measured relative to the S&P 500 as the benchmark.
+             Use this when the user asks how to allocate investments to minimize underperformance
+             relative to the market.
+             Pass each ticker as a separate argument — never as a single string.
+Returns: "Optimal weights for minimum target semivariance portfolio:
+          {TICKER: weight, ...}
+          Expected annual return: <return>
+          Annualized volatility: <volatility>"
+Example call: ACTION: min_target_semivariance_portfolio(AAPL, MSFT, GOOGL)
+
+### get_cetes_28(date)
+Description: Returns the CETES 28-day yield from Banxico.
+             If no date is provided, returns the most recent observation.
+             Pass a date in YYYY-MM-DD format to get the nearest available observation.
+Returns: "The CETES 28-day rate (<label>) is <value>% as of <DD/MM/YYYY>."
+Example calls: ACTION: get_cetes_28()
+               ACTION: get_cetes_28(2024-01-15)
+
+### get_cetes_91(date)
+Description: Returns the CETES 91-day yield from Banxico.
+             If no date is provided, returns the most recent observation.
+Returns: "The CETES 91-day rate (<label>) is <value>% as of <DD/MM/YYYY>."
+Example calls: ACTION: get_cetes_91()
+               ACTION: get_cetes_91(2024-01-15)
+
+### get_cetes_182(date)
+Description: Returns the CETES 182-day yield from Banxico.
+             If no date is provided, returns the most recent observation.
+Returns: "The CETES 182-day rate (<label>) is <value>% as of <DD/MM/YYYY>."
+Example calls: ACTION: get_cetes_182()
+               ACTION: get_cetes_182(2024-01-15)
+
+### get_cetes_364(date)
+Description: Returns the CETES 364-day yield from Banxico.
+             If no date is provided, returns the most recent observation.
+Returns: "The CETES 364-day rate (<label>) is <value>% as of <DD/MM/YYYY>."
+Example calls: ACTION: get_cetes_364()
+               ACTION: get_cetes_364(2024-01-15)
+
+### get_cetes_728(date)
+Description: Returns the CETES 728-day yield from Banxico.
+             If no date is provided, returns the most recent observation.
+Returns: "The CETES 728-day rate (<label>) is <value>% as of <DD/MM/YYYY>."
+Example calls: ACTION: get_cetes_728()
+               ACTION: get_cetes_728(2024-01-15)
+
+### get_tie_28(date)
+Description: Returns the TIE (Tasa de Interés de Equilibrio) 28-day rate from Banxico.
+             If no date is provided, returns the most recent observation.
+Returns: "The TIE 28-day rate (<label>) is <value>% as of <DD/MM/YYYY>."
+Example calls: ACTION: get_tie_28()
+               ACTION: get_tie_28(2024-01-15)
+
+### get_tie_91(date)
+Description: Returns the TIE 91-day rate from Banxico.
+             If no date is provided, returns the most recent observation.
+Returns: "The TIE 91-day rate (<label>) is <value>% as of <DD/MM/YYYY>."
+Example calls: ACTION: get_tie_91()
+               ACTION: get_tie_91(2024-01-15)
+
+### get_tie_182(date)
+Description: Returns the TIE 182-day rate from Banxico.
+             If no date is provided, returns the most recent observation.
+Returns: "The TIE 182-day rate (<label>) is <value>% as of <DD/MM/YYYY>."
+Example calls: ACTION: get_tie_182()
+               ACTION: get_tie_182(2024-01-15)
+
+### get_target_interest_rate_mexico(date)
+Description: Returns the Banxico target interest rate (tasa objetivo).
+             If no date is provided, returns the most recent observation.
+Returns: "The target interest rate in Mexico (<label>) is <value>% as of <DD/MM/YYYY>."
+Example calls: ACTION: get_target_interest_rate_mexico()
+               ACTION: get_target_interest_rate_mexico(2024-06-01)
+
+### get_mensual_inflation_mexico(date)
+Description: Returns the monthly inflation rate in Mexico from Banxico.
+             If no date is provided, returns the most recent observation.
+Returns: "The monthly inflation rate in Mexico (<label>) is <value>% as of <DD/MM/YYYY>."
+Example calls: ACTION: get_mensual_inflation_mexico()
+               ACTION: get_mensual_inflation_mexico(2024-06-01)
+
+### get_inflation_mexico(date)
+Description: Returns the annual inflation rate in Mexico from Banxico.
+             If no date is provided, returns the most recent observation.
+Returns: "The annual inflation rate in Mexico (<label>) is <value>% as of <DD/MM/YYYY>."
+Example calls: ACTION: get_inflation_mexico()
+               ACTION: get_inflation_mexico(2024-06-01)
+
+### get_udis(date)
+Description: Returns the value of UDIs (Unidades de Inversión) in MXN from Banxico.
+             If no date is provided, returns the most recent observation.
+Returns: "The value of UDIs in Mexico (<label>) is <value> MXN as of <DD/MM/YYYY>."
+Example calls: ACTION: get_udis()
+               ACTION: get_udis(2024-06-01)
+
+### get_exchange_rate(base, quote, date)
+Description: Returns the market exchange rate between any two currencies using yfinance.
+             base and quote must be ISO 4217 codes (e.g. 'EUR', 'USD', 'GBP').
+             If no date is provided, returns the most recent available rate.
+             Pass a date in YYYY-MM-DD format to get the nearest available rate.
+Returns: "The exchange rate for <BASE>/<QUOTE> is <VALUE> (<label>) as of <YYYY-MM-DD>."
+Example calls: ACTION: get_exchange_rate(EUR, USD)
+               ACTION: get_exchange_rate(GBP, JPY, 2024-03-15)
+               ACTION: get_exchange_rate(USD, CAD)
+
+### get_news_sentiment(ticker)
+Description: Fetches recent news articles for a stock ticker and returns a FinBERT-based
+             sentiment score aggregated across all available headlines.
+             Each article is scored (positive / neutral / negative) and weighted by recency
+             using exponential decay so the most recent news has the highest influence.
+             The composite score ranges from -1 (fully negative) to +1 (fully positive);
+             scores above 0.15 are labelled POSITIVE, below -0.15 are NEGATIVE, otherwise NEUTRAL.
+             Apply the same exchange suffix rules as get_price_on_date.
+             Use this when the user asks about market sentiment, news tone, or recent coverage of a stock.
+Returns: "Sentiment analysis for <Company Name> (<TICKER>) across <N> recent articles:
+          Composite score: <score> (<LABEL>).
+          Top influencing headlines: [LABEL CONFIDENCE%] <title> (<provider>) --- ..."
+Example calls: ACTION: get_news_sentiment(AAPL)
+               ACTION: get_news_sentiment(TSLA)
+               ACTION: get_news_sentiment(BIMBOA.MX)
+
+### respond_to_greeting()
+Description: Responds to user greetings with a friendly introduction about the agent.
+             Use this when the user greets you or asks a general question like "Hi", "Hello", "What are you?".
+Returns: "Hello! I'm a financial data agent. How can I assist you today?"
+Example call: ACTION: respond_to_greeting()
+
+### respond_no_available_tool(tool_name)
+Description: Responds when a user asks for a tool or action that is not available.
+             Use this when the user requests functionality that doesn't match any available tool.
+Returns: "Sorry, currently i'm capable of doing that. Check the list of available tools with 'list_tools' command."
+Example call: ACTION: respond_no_available_tool()
+
+=== END OF TOOLS ===
+
+EXAMPLES:
+
+User: What is the price of Microsoft?
+ACTION: get_price_on_date(MSFT)
+After tool result: FINAL: The last price of Microsoft Corporation (MSFT) is $415.20 as of 2026-04-29.
+
+User: What was Apple's price on March 10 2023?
+ACTION: get_price_on_date(AAPL, 2023-03-10)
+After tool result: FINAL: Apple's closing price nearest to March 10, 2023 was $150.02 on 2023-03-10.
+
+User: What is the price of Bimbo?
+ACTION: get_price_on_date(BIMBOA.MX)
+After tool result: FINAL: The last price of Grupo Bimbo S.A.B. de C.V. (BIMBOA.MX) is $X.XX as of 2026-04-29.
+
+User: What was the price of Jose Cuervo on March 1 2024?
+ACTION: get_price_on_date(CUERVO.MX, 2024-03-01)
+After tool result: FINAL: The price of Jose Cuervo Internacional S.A.B. de C.V. (CUERVO.MX) nearest to March 1, 2024 was $X.XX on 2024-03-01.
+
+User: What does Grupo México do?
+ACTION: get_company_profile(GMEXICOB.MX)
+After tool result: FINAL: Grupo México operates in the Basic Materials sector and Copper industry. The company is one of the largest mining groups in Latin America, focused on copper, silver, and zinc extraction.
+
+
+ACTION: get_company_profile(NVDA)
+After tool result: FINAL: Nvidia operates in the Technology sector and Semiconductors industry. The company designs GPUs and accelerated computing platforms for gaming, data centers, and artificial intelligence.
+
+User: How should I allocate $100k between Apple, Nvidia, and Intel to minimize risk?
+ACTION: min_variance_portfolio(AAPL, NVDA, INTC)
+After tool result: FINAL: To minimize risk, the optimal weights across Apple, Nvidia, and Intel are AAPL: 45%, NVDA: 35%, INTC: 20%, with an expected annual return of 18.40% and annualized volatility of 22.10%.
+
+User: How should I allocate $100k between Apple, Nvidia, and Intel to maximize risk-adjusted return?
+ACTION: max_sharpe_portfolio(AAPL, NVDA, INTC)
+After tool result: FINAL: To maximize risk-adjusted return, the optimal weights across Apple, Nvidia, and Intel are AAPL: 30%, NVDA: 50%, INTC: 20%, with an expected annual return of 20.20% and annualized volatility of 23.50%.
+
+User: How should I allocate $100k between Apple, Nvidia, and Intel to minimize downside risk below the S&P 500?
+ACTION: min_target_semivariance_portfolio(AAPL, NVDA, INTC)
+After tool result: FINAL: To minimize downside risk relative to the S&P 500, the optimal weights across Apple, Nvidia, and Intel are AAPL: 40%, NVDA: 30%, INTC: 30%, with an expected annual return of 19.00% and annualized volatility of 15.80%.
+
+User: What is the sentiment around Tesla news?
+ACTION: get_news_sentiment(TSLA)
+After tool result: FINAL: Sentiment analysis for Tesla, Inc. (TSLA) across 8 recent articles: Composite score: +0.3142 (POSITIVE). Top influencing headlines: [POSITIVE 92%] Tesla beats Q1 delivery estimates (Reuters) --- [NEGATIVE 85%] Tesla faces renewed pressure on margins (Bloomberg) --- ...
+
+User: What is the market sentiment for Apple stock right now?
+ACTION: get_news_sentiment(AAPL)
+After tool result: FINAL: Sentiment analysis for Apple Inc. (AAPL) across 10 recent articles: Composite score: +0.1823 (POSITIVE). Top influencing headlines: [POSITIVE 88%] Apple reports record services revenue (CNBC) --- ...
+
+User: What is the most recent CETES 28-day rate?
+ACTION: get_cetes_28()
+After tool result: FINAL: The CETES 28-day rate (most recent) is 8.9900% as of 27/03/2025.
+
+User: What was the CETES 91-day rate in January 2024?
+ACTION: get_cetes_91(2024-01-15)
+After tool result: FINAL: The CETES 91-day rate nearest to January 15, 2024 was 11.3100% as of 11/01/2024.
+
+User: What is the current Banxico target rate?
+ACTION: get_target_interest_rate_mexico()
+After tool result: FINAL: The Banxico target interest rate (most recent) is 9.0000% as of 20/03/2025.
+
+User: What is the annual inflation rate in Mexico?
+ACTION: get_inflation_mexico()
+After tool result: FINAL: The annual inflation rate in Mexico (most recent) is 3.8000% as of 28/02/2025.
+
+User: What was the monthly inflation in Mexico in mid-2023?
+ACTION: get_mensual_inflation_mexico(2023-06-15)
+After tool result: FINAL: The monthly inflation rate in Mexico nearest to June 15, 2023 was 0.2200% as of 15/06/2023.
+
+User: What is the TIE 182-day rate?
+ACTION: get_tie_182()
+After tool result: FINAL: The TIE 182-day rate (most recent) is 9.4500% as of 15/04/2025.
+
+User: What is the current UDI value?
+ACTION: get_udis()
+After tool result: FINAL: The value of UDIs in Mexico (most recent) is 8.2341 MXN as of 30/04/2025.
+
+User: Hi there!
+ACTION: respond_to_greeting()
+After tool result: FINAL: Hello! I'm a financial data agent. How can I assist you today?
+
+User: Can you tell me how many calories are in an apple?
+ACTION: respond_no_available_tool()
+After tool result: FINAL: Sorry, currently i'm capable of doing that. Check the list of avaiable tools for more information.
+
+INCORRECT (never do this):
+ACTION: get_price_on_date(BIMBOA)                    <- missing .MX suffix for BMV-listed stock
+ACTION: get_price_on_date(CUERVO)                    <- missing .MX suffix; correct is CUERVO.MX
+ACTION: get_last_price(AAPL)                         <- wrong tool name; the correct name is get_price_on_date
+ACTION: get_price_on_date("Apple")                   <- use ticker symbol, not a company name string
+ACTION: get_price_on_date(AAPL, MSFT)                <- second argument must be a date, not another ticker
+ACTION: get_price_on_date(AAPL, March 2023)          <- date must be in YYYY-MM-DD format
+ACTION: min_variance_portfolio("AAPL, MSFT, GOOGL")  <- never pack tickers into one string argument
+FINAL: Apple's price is around $210                  <- invented value, no tool was called
+FINAL: The CETES rate is roughly 9%                  <- recalled from memory, no tool was called
+"""

requirements.txt

@@ -0,0 +1,10 @@
+gradio==6.13.0
+openai==2.33.0
+yfinance==1.3.0
+numpy==2.4.4
+pandas==3.0.2
+scipy==1.17.1
+requests==2.33.1
+transformers>=4.40.0
+torch>=2.2.0
+huggingface_hub>=0.20.0

tools.py

@@ -0,0 +1,739 @@
+from dataclasses import dataclass, field
+from typing import Callable
+
+import numpy as np
+from scipy.optimize import minimize
+import yfinance as yf
+import requests
+import datetime
+import pandas as pd
+
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import torch
+import torch.nn.functional as F
+
+import os
+
+BANXICO_TOKEN = os.getenv("BANXICO_TOKEN")
+HF_LOGIN_KEY = os.getenv("HF_LOGIN_KEY")
+if HF_LOGIN_KEY:
+    from huggingface_hub import login
+    login(HF_LOGIN_KEY)
+
+ToolFunction = Callable[..., object]
+
+_DEFAULT_TOOL_FUNCTIONS: dict[str, ToolFunction] = {}
+
+
+def tool(name: str | None = None):
+    """Decorator that registers a function in the default tool registry."""
+    def decorator(function: ToolFunction) -> ToolFunction:
+        _DEFAULT_TOOL_FUNCTIONS[name or function.__name__] = function
+        return function
+    return decorator
+
+
+@dataclass
+class ToolRegistry:
+    tools: dict[str, ToolFunction] = field(default_factory=dict)
+
+    def register(self, name: str, function: ToolFunction) -> None:
+        self.tools[name] = function
+
+    def execute(self, name: str, *args) -> object:
+        if name not in self.tools:
+            raise KeyError(f"tool '{name}' does not exist")
+        return self.tools[name](*args)
+
+    def names(self) -> list[str]:
+        return list(self.tools.keys())
+
+
+def build_default_tool_registry() -> ToolRegistry:
+    registry = ToolRegistry()
+    for name, function in _DEFAULT_TOOL_FUNCTIONS.items():
+        registry.register(name, function)
+    return registry
+
+
+@tool("get_price_on_date")
+def get_price_on_date(ticker, date=None):
+    t = yf.Ticker(ticker)
+
+    use_default_date = date is None
+
+    if date is None:
+        date = datetime.date.today()
+    else:
+        date = datetime.datetime.strptime(date, "%Y-%m-%d").date()
+
+    data = pd.DataFrame(t.history(start=date - datetime.timedelta(days=5), end=date + datetime.timedelta(days=5))['Close'])
+    if data.empty:
+        return f"No price data available for {t.ticker} around {date}."
+
+    data['Date'] = data.index.date
+    data['DateDiff'] = np.abs(data['Date'] - date)
+    nearest_row = data.loc[data['DateDiff'].idxmin()]
+    price = nearest_row['Close']
+    actual_date = nearest_row['Date']
+    official_name = t.info['longName']
+
+    if use_default_date:
+        return f"The last price of {official_name} ({t.ticker}) is ${price:.2f} as of {actual_date}."
+    else:
+        return f"The price of {official_name} ({t.ticker}) nearest to {date} was ${price:.2f} on {actual_date}."
+
+@tool("get_company_profile")
+def get_company_profile_tool(ticker: str) -> str:
+    t = yf.Ticker(ticker)
+    info = t.info
+    official_name = info['longName']
+    sector = info.get('sector', 'N/A')
+    industry = info.get('industry', 'N/A')
+    description = info.get('longBusinessSummary', 'No description available.')
+    return (
+        f"{official_name} operates in the {sector} sector and {industry} industry. "
+        f"Company profile: {description}"
+    )
+
+@tool("min_variance_portfolio")
+def min_variance_portfolio(*tickers: str) -> str:
+    ticker_list = list(tickers)
+    data = yf.download(ticker_list, period="2y", progress=False)['Close'][ticker_list]
+    returns = data.pct_change().dropna()
+    cov_matrix = returns.cov()
+    mean_rt = returns.mean()
+
+    variance = lambda w: w.T @ cov_matrix @ w
+    x0 = np.ones(len(ticker_list)) / len(ticker_list)
+    bounds = [(0, 3)] * len(ticker_list)
+    constraints = {'type': 'eq', 'fun': lambda w: np.sum(w) - 1}
+    result = minimize(variance, x0, bounds=bounds, constraints=constraints, tol=1e-16, method='SLSQP')
+    return (
+        f"Optimal weights for minimum variance portfolio:\n"
+        f"{ {ticker_list[i]: round(w, 4) for i, w in enumerate(result.x)} }\n"
+        f"Expected annual return: {(mean_rt @ result.x * 252):.2%}\n"
+        f"Annualized volatility: {(np.sqrt(result.x.T @ cov_matrix @ result.x) * np.sqrt(252)):.2%}"
+    )
+
+@tool("max_sharpe_portfolio")
+def max_sharpe_portfolio(*tickers: str) -> str:
+    ticker_list = list(tickers)
+    data = yf.download(ticker_list, period="2y", progress=False)['Close'][ticker_list]
+    returns = data.pct_change().dropna()
+    cov_matrix = returns.cov()
+    mean_rt = returns.mean()
+
+    sharpe = lambda w: -(mean_rt @ w) / np.sqrt(w.T @ cov_matrix @ w)
+    x0 = np.ones(len(ticker_list)) / len(ticker_list)
+    bounds = [(0, 3)] * len(ticker_list)
+    constraints = {'type': 'eq', 'fun': lambda w: np.sum(w) - 1}
+    result = minimize(sharpe, x0, bounds=bounds, constraints=constraints, tol=1e-16, method='SLSQP')
+    return (
+        f"Optimal weights for maximum Sharpe ratio portfolio:\n"
+        f"{ {ticker_list[i]: round(w, 4) for i, w in enumerate(result.x)} }\n"
+        f"Expected annual return: {(mean_rt @ result.x * 252):.2%}\n"
+        f"Annualized volatility: {(np.sqrt(result.x.T @ cov_matrix @ result.x) * np.sqrt(252)):.2%}"
+    )
+
+@tool("min_target_semivariance_portfolio")
+def min_target_semivariance_portfolio(*tickers: str) -> str:
+    ticker_list = list(tickers)
+    data = yf.download(ticker_list, period="2y", progress=False)['Close'][ticker_list]
+    returns = data.pct_change().dropna()
+    corr = returns.corr()
+    cov_matrix = returns.cov()
+    benchmark = yf.download("^GSPC", period="2y", progress=False)['Close'].pct_change().dropna()
+    differences = returns - benchmark.values
+    below_zero_target = differences[differences < 0].fillna(0)
+    target_downside = np.array(below_zero_target.std())
+    target_semivariance = np.multiply(target_downside.reshape(len(target_downside), 1), target_downside) * corr
+
+    semivariance = lambda w: w.T @ target_semivariance @ w
+    x0 = np.ones(len(ticker_list)) / len(ticker_list)
+    bounds = [(0, 3)] * len(ticker_list)
+    constraints = {'type': 'eq', 'fun': lambda w: np.sum(w) - 1}
+    result = minimize(semivariance, x0, bounds=bounds, constraints=constraints, tol=1e-16, method='SLSQP')
+    return (
+        f"Optimal weights for minimum target semivariance portfolio:\n"
+        f"{ {ticker_list[i]: round(w, 4) for i, w in enumerate(result.x)} }\n"
+        f"Expected annual return: {(returns.mean() @ result.x * 252):.2%}\n"
+        f"Annualized volatility: {(np.sqrt(result.x.T @ cov_matrix @ result.x) * np.sqrt(252)):.2%}"
+    )
+
+@tool("get_cetes_28")
+def get_cetes_28(date: str | None = None) -> str:
+    URL = "https://www.banxico.org.mx/SieAPIRest/service/v1/series/SF43936/datos"
+    headers = {
+        "Bmx-Token": BANXICO_TOKEN,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.get(URL, headers=headers)
+        response.raise_for_status()
+
+        obs_list = response.json()["bmx"]["series"][0]["datos"]
+
+        if date is None:
+            obs = obs_list[-1]
+        else:
+            target = datetime.datetime.strptime(date, "%Y-%m-%d")
+            obs = min(
+                obs_list,
+                key=lambda o: abs(datetime.datetime.strptime(o["fecha"], "%d/%m/%Y") - target),
+            )
+
+        fecha = obs["fecha"]
+        fecha = datetime.datetime.strptime(fecha, "%d/%m/%Y").strftime("%Y-%m-%d")
+        value = float(obs["dato"])
+        label = f"nearest to {date}" if date else "most recent"
+        return f"The CETES 28-day rate ({label}) is {value:.4f}% as of {fecha}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching CETES 28-day rate: {exc}"
+
+@tool("get_cetes_91")
+def get_cetes_91(date: str | None = None) -> str:
+    URL = "https://www.banxico.org.mx/SieAPIRest/service/v1/series/SF43939/datos"
+    headers = {
+        "Bmx-Token": BANXICO_TOKEN,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.get(URL, headers=headers)
+        response.raise_for_status()
+
+        obs_list = response.json()["bmx"]["series"][0]["datos"]
+
+        if date is None:
+            obs = obs_list[-1]
+        else:
+            target = datetime.datetime.strptime(date, "%Y-%m-%d")
+            obs = min(
+                obs_list,
+                key=lambda o: abs(datetime.datetime.strptime(o["fecha"], "%d/%m/%Y") - target),
+            )
+
+        fecha = obs["fecha"]
+        fecha = datetime.datetime.strptime(fecha, "%d/%m/%Y").strftime("%Y-%m-%d")
+        value = float(obs["dato"])
+        label = f"nearest to {date}" if date else "most recent"
+        return f"The CETES 91-day rate ({label}) is {value:.4f}% as of {fecha}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching CETES 91-day rate: {exc}"
+
+@tool("get_cetes_182")
+def get_cetes_182(date: str | None = None) -> str:
+    URL = "https://www.banxico.org.mx/SieAPIRest/service/v1/series/SF43942/datos"
+    headers = {
+        "Bmx-Token": BANXICO_TOKEN,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.get(URL, headers=headers)
+        response.raise_for_status()
+
+        obs_list = response.json()["bmx"]["series"][0]["datos"]
+
+        if date is None:
+            obs = obs_list[-1]
+        else:
+            target = datetime.datetime.strptime(date, "%Y-%m-%d")
+            obs = min(
+                obs_list,
+                key=lambda o: abs(datetime.datetime.strptime(o["fecha"], "%d/%m/%Y") - target),
+            )
+
+        fecha = obs["fecha"]
+        fecha = datetime.datetime.strptime(fecha, "%d/%m/%Y").strftime("%Y-%m-%d")
+        value = float(obs["dato"])
+        label = f"nearest to {date}" if date else "most recent"
+        return f"The CETES 182-day rate ({label}) is {value:.4f}% as of {fecha}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching CETES 182-day rate: {exc}"
+
+@tool("get_cetes_364")
+def get_cetes_364(date: str | None = None) -> str:
+    URL = "https://www.banxico.org.mx/SieAPIRest/service/v1/series/SF43945/datos"
+    headers = {
+        "Bmx-Token": BANXICO_TOKEN,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.get(URL, headers=headers)
+        response.raise_for_status()
+
+        obs_list = response.json()["bmx"]["series"][0]["datos"]
+
+        if date is None:
+            obs = obs_list[-1]
+        else:
+            target = datetime.datetime.strptime(date, "%Y-%m-%d")
+            obs = min(
+                obs_list,
+                key=lambda o: abs(datetime.datetime.strptime(o["fecha"], "%d/%m/%Y") - target),
+            )
+
+        fecha = obs["fecha"]
+        fecha = datetime.datetime.strptime(fecha, "%d/%m/%Y").strftime("%Y-%m-%d")
+        value = float(obs["dato"])
+        label = f"nearest to {date}" if date else "most recent"
+        return f"The CETES 364-day rate ({label}) is {value:.4f}% as of {fecha}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching CETES 364-day rate: {exc}"
+
+@tool("get_cetes_728")
+def get_cetes_728(date: str | None = None) -> str:
+    URL = "https://www.banxico.org.mx/SieAPIRest/service/v1/series/SF349785/datos"
+    headers = {
+        "Bmx-Token": BANXICO_TOKEN,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.get(URL, headers=headers)
+        response.raise_for_status()
+
+        obs_list = response.json()["bmx"]["series"][0]["datos"]
+
+        if date is None:
+            obs = obs_list[-1]
+        else:
+            target = datetime.datetime.strptime(date, "%Y-%m-%d")
+            obs = min(
+                obs_list,
+                key=lambda o: abs(datetime.datetime.strptime(o["fecha"], "%d/%m/%Y") - target),
+            )
+
+        fecha = obs["fecha"]
+        fecha = datetime.datetime.strptime(fecha, "%d/%m/%Y").strftime("%Y-%m-%d")
+        value = float(obs["dato"])
+        label = f"nearest to {date}" if date else "most recent"
+        return f"The CETES 728-day rate ({label}) is {value:.4f}% as of {fecha}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching CETES 728-day rate: {exc}"
+
+@tool("get_mensual_inflation_mexico")
+def get_mensual_inflation_mexico(date: str | None = None) -> str:
+    URL = "https://www.banxico.org.mx/SieAPIRest/service/v1/series/SP30578/datos"
+    headers = {
+        "Bmx-Token": BANXICO_TOKEN,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.get(URL, headers=headers)
+        response.raise_for_status()
+
+        obs_list = response.json()["bmx"]["series"][0]["datos"]
+
+        if date is None:
+            obs = obs_list[-1]
+        else:
+            target = datetime.datetime.strptime(date, "%Y-%m-%d")
+            obs = min(
+                obs_list,
+                key=lambda o: abs(datetime.datetime.strptime(o["fecha"], "%d/%m/%Y") - target),
+            )
+
+        fecha = obs["fecha"]
+        fecha = datetime.datetime.strptime(fecha, "%d/%m/%Y").strftime("%Y-%m-%d")
+        value = float(obs["dato"])
+        label = f"nearest to {date}" if date else "most recent"
+        return f"The monthly inflation rate in Mexico ({label}) is {value:.4f}% as of {fecha}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching monthly inflation rate in Mexico: {exc}"
+
+@tool("get_inflation_mexico")
+def get_inflation_mexico(date: str | None = None) -> str:
+    URL = "https://www.banxico.org.mx/SieAPIRest/service/v1/series/SP30577/datos"
+    headers = {
+        "Bmx-Token": BANXICO_TOKEN,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.get(URL, headers=headers)
+        response.raise_for_status()
+
+        obs_list = response.json()["bmx"]["series"][0]["datos"]
+
+        if date is None:
+            obs = obs_list[-1]
+        else:
+            target = datetime.datetime.strptime(date, "%Y-%m-%d")
+            obs = min(
+                obs_list,
+                key=lambda o: abs(datetime.datetime.strptime(o["fecha"], "%d/%m/%Y") - target),
+            )
+
+        fecha = obs["fecha"]
+        fecha = datetime.datetime.strptime(fecha, "%d/%m/%Y").strftime("%Y-%m-%d")
+        value = float(obs["dato"])
+        label = f"nearest to {date}" if date else "most recent"
+        return f"The annual inflation rate in Mexico ({label}) is {value:.4f}% as of {fecha}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching annual inflation rate in Mexico: {exc}"
+
+@tool("get_udis")
+def get_udis(date: str | None = None) -> str:
+    URL = "https://www.banxico.org.mx/SieAPIRest/service/v1/series/SP68257/datos"
+    headers = {
+        "Bmx-Token": BANXICO_TOKEN,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.get(URL, headers=headers)
+        response.raise_for_status()
+
+        obs_list = response.json()["bmx"]["series"][0]["datos"]
+
+        if date is None:
+            obs = obs_list[-1]
+        else:
+            target = datetime.datetime.strptime(date, "%Y-%m-%d")
+            obs = min(
+                obs_list,
+                key=lambda o: abs(datetime.datetime.strptime(o["fecha"], "%d/%m/%Y") - target),
+            )
+
+        fecha = obs["fecha"]
+        fecha = datetime.datetime.strptime(fecha, "%d/%m/%Y").strftime("%Y-%m-%d")
+        value = float(obs["dato"])
+        label = f"nearest to {date}" if date else "most recent"
+        return f"The value of UDIs in Mexico ({label}) is {value:.4f} MXN as of {fecha}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching UDIs value in Mexico: {exc}"
+
+@tool("get_tie_28")
+def get_tie_28(date: str | None = None) -> str:
+    URL = "https://www.banxico.org.mx/SieAPIRest/service/v1/series/SF43783/datos"
+    headers = {
+        "Bmx-Token": BANXICO_TOKEN,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.get(URL, headers=headers)
+        response.raise_for_status()
+
+        obs_list = response.json()["bmx"]["series"][0]["datos"]
+
+        if date is None:
+            obs = obs_list[-1]
+        else:
+            target = datetime.datetime.strptime(date, "%Y-%m-%d")
+            obs = min(
+                obs_list,
+                key=lambda o: abs(datetime.datetime.strptime(o["fecha"], "%d/%m/%Y") - target),
+            )
+
+        fecha = obs["fecha"]
+        fecha = datetime.datetime.strptime(fecha, "%d/%m/%Y").strftime("%Y-%m-%d")
+        value = float(obs["dato"])
+        label = f"nearest to {date}" if date else "most recent"
+        return f"The TIE 28-day rate ({label}) is {value:.4f}% as of {fecha}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching TIE 28-day rate: {exc}"
+
+@tool("get_tie_91")
+def get_tie_91(date: str | None = None) -> str:
+    URL = "https://www.banxico.org.mx/SieAPIRest/service/v1/series/SF43878/datos"
+    headers = {
+        "Bmx-Token": BANXICO_TOKEN,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.get(URL, headers=headers)
+        response.raise_for_status()
+
+        obs_list = response.json()["bmx"]["series"][0]["datos"]
+
+        if date is None:
+            obs = obs_list[-1]
+        else:
+            target = datetime.datetime.strptime(date, "%Y-%m-%d")
+            obs = min(
+                obs_list,
+                key=lambda o: abs(datetime.datetime.strptime(o["fecha"], "%d/%m/%Y") - target),
+            )
+
+        fecha = obs["fecha"]
+        fecha = datetime.datetime.strptime(fecha, "%d/%m/%Y").strftime("%Y-%m-%d")
+        value = float(obs["dato"])
+        label = f"nearest to {date}" if date else "most recent"
+        return f"The TIE 91-day rate ({label}) is {value:.4f}% as of {fecha}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching TIE 91-day rate: {exc}"
+
+@tool("get_tie_182")
+def get_tie_182(date: str | None = None) -> str:
+    URL = "https://www.banxico.org.mx/SieAPIRest/service/v1/series/SF111916/datos"
+    headers = {
+        "Bmx-Token": BANXICO_TOKEN,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.get(URL, headers=headers)
+        response.raise_for_status()
+
+        obs_list = response.json()["bmx"]["series"][0]["datos"]
+
+        if date is None:
+            obs = obs_list[-1]
+        else:
+            target = datetime.datetime.strptime(date, "%Y-%m-%d")
+            obs = min(
+                obs_list,
+                key=lambda o: abs(datetime.datetime.strptime(o["fecha"], "%d/%m/%Y") - target),
+            )
+
+        fecha = obs["fecha"]
+        fecha = datetime.datetime.strptime(fecha, "%d/%m/%Y").strftime("%Y-%m-%d")
+        value = float(obs["dato"])
+        label = f"nearest to {date}" if date else "most recent"
+        return f"The TIE 182-day rate ({label}) is {value:.4f}% as of {fecha}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching TIE 182-day rate: {exc}"
+
+@tool("get_target_interest_rate_mexico")
+def get_target_interest_rate_mexico(date: str | None = None) -> str:
+    URL = "https://www.banxico.org.mx/SieAPIRest/service/v1/series/SF61745/datos"
+    headers = {
+        "Bmx-Token": BANXICO_TOKEN,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.get(URL, headers=headers)
+        response.raise_for_status()
+
+        obs_list = response.json()["bmx"]["series"][0]["datos"]
+
+        if date is None:
+            obs = obs_list[-1]
+        else:
+            target = datetime.datetime.strptime(date, "%Y-%m-%d")
+            obs = min(
+                obs_list,
+                key=lambda o: abs(datetime.datetime.strptime(o["fecha"], "%d/%m/%Y") - target),
+            )
+
+        fecha = obs["fecha"]
+        fecha = datetime.datetime.strptime(fecha, "%d/%m/%Y").strftime("%Y-%m-%d")
+        value = float(obs["dato"])
+        label = f"nearest to {date}" if date else "most recent"
+        return f"The target interest rate in Mexico ({label}) is {value:.4f}% as of {fecha}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching target interest rate in Mexico: {exc}"
+
+@tool("get_exchange_rate")
+def get_exchange_rate(base: str, quote: str, date: str | None = None) -> str:
+    base = base.strip().upper()
+    quote = quote.strip().upper()
+    ticker_symbol = f"{base}{quote}=X"
+
+    try:
+        if date is None:
+            target_date = datetime.date.today()
+        else:
+            target_date = datetime.datetime.strptime(date, "%Y-%m-%d").date()
+
+        t = yf.Ticker(ticker_symbol)
+        data = t.history(
+            start=target_date - datetime.timedelta(days=7),
+            end=target_date + datetime.timedelta(days=7),
+        )
+
+        if data.empty:
+            return (
+                f"No exchange rate data found for {base}/{quote} ({ticker_symbol}). "
+                f"Verify that both currency codes are valid ISO 4217 codes."
+            )
+
+        data["Date"] = data.index.date
+        data["DateDiff"] = data["Date"].apply(lambda d: abs((d - target_date).days))
+        nearest = data.loc[data["DateDiff"].idxmin()]
+        rate = nearest["Close"]
+        actual_date = nearest["Date"]
+        date_label = f"nearest to {date}" if date else "most recent"
+        return f"The exchange rate for {base}/{quote} ({date_label}) is {rate:.6f} as of {actual_date}."
+
+    except ValueError:
+        return f"Invalid date format '{date}'. Please use YYYY-MM-DD (e.g. 2024-01-15)."
+    except Exception as exc:
+        return f"Error fetching exchange rate for {base}/{quote}: {exc}"
+
+def _get_news(ticker: str) -> list[dict]:
+    t = yf.Ticker(ticker)
+    news = t.news
+    formated_news = []
+
+    for i in range(len(news)):
+        item = news[i]['content']
+        formated_news.append({
+            "pub_date": item.get("pubDate", ""),
+            "content_type": item.get("contentType", ""),
+            "title": item.get("title", ""),
+            "summary": item.get("summary", ""),
+            "provider": item.get("provider", {}).get("displayName", "N/A"),
+        })
+    return formated_news
+
+_finbert_tokenizer = None
+_finbert_model = None
+
+def _load_finbert():
+    global _finbert_tokenizer, _finbert_model
+    if _finbert_model is None:
+        _finbert_tokenizer = AutoTokenizer.from_pretrained("ProsusAI/finbert")
+        _finbert_model = AutoModelForSequenceClassification.from_pretrained("ProsusAI/finbert")
+        _finbert_model.eval()
+    return _finbert_tokenizer, _finbert_model
+
+_LABEL_TO_SCORE = {
+    "positive": 1,
+    "neutral": 0,
+    "negative": -1
+}
+
+def _bucket_label(score: float) -> str:
+    if score > 0.15:
+        return "positive"
+    if score < -0.15:
+        return "negative"
+    return "neutral"
+
+def _recency_weights(pub_dates: list[str]) -> list[float]:
+    decay = 0.01
+    parsed = []
+    for d in pub_dates:
+        try:
+            dt = datetime.datetime.fromisoformat(d.replace("Z", "+00:00"))
+            parsed.append(dt)
+        except (ValueError, AttributeError):
+            parsed.append(None)
+
+    valid = [dt for dt in parsed if dt is not None]
+    if not valid:
+        return [1.0] * len(pub_dates)
+
+    most_recent = max(valid)
+    weights = []
+    for dt in parsed:
+        if dt is None:
+            weights.append(0.5)
+        else:
+            hours_old = (most_recent - dt).total_seconds() / 3600
+            weights.append(float(np.exp(-decay * hours_old)))
+    return weights
+
+def _score_texts(texts: list[str]) -> list[dict]:
+    tokenizer, model = _load_finbert()
+    results = []
+    with torch.no_grad():
+        for text in texts:
+            inputs = tokenizer(
+                text,
+                return_tensors="pt",
+                truncation=True,
+                max_length=512,
+                padding=True,
+            )
+            logits = model(**inputs).logits
+            probs  = F.softmax(logits, dim=-1).squeeze()
+            # FinBERT label order: positive=0, negative=1, neutral=2
+            label_map = {0: "positive", 1: "negative", 2: "neutral"}
+            pred_idx   = int(probs.argmax())
+            results.append({
+                "label":      label_map[pred_idx],
+                "confidence": float(probs[pred_idx]),
+            })
+    return results
+
+@tool("get_news_sentiment")
+def get_news_sentiment(ticker: str) -> str:
+    articles = _get_news(ticker)
+    comp_name = yf.Ticker(ticker).info.get("longName", ticker)
+    if not articles:
+        return f"No recent news found for {ticker}."
+
+    texts   = [f"{a['title']}. {a['summary']}".strip() for a in articles]
+    scores  = _score_texts(texts)
+    weights = _recency_weights([a["pub_date"] for a in articles])
+
+    weighted_sum   = 0.0
+    total_weight   = 0.0
+    scored_articles = []
+
+    for article, score, weight in zip(articles, scores, weights):
+        numeric = _LABEL_TO_SCORE[score["label"]]
+        contribution = numeric * score["confidence"] * weight
+        weighted_sum  += contribution
+        total_weight  += weight
+        scored_articles.append({
+            "title":       article["title"],
+            "provider":    article["provider"],
+            "label":       score["label"],
+            "confidence":  score["confidence"],
+            "weight":      round(weight, 4),
+            "pub_date":    article["pub_date"],
+        })
+
+    composite = weighted_sum / total_weight if total_weight > 0 else 0.0
+    composite = max(-1.0, min(1.0, composite))
+    label     = _bucket_label(composite)
+    label     = label.upper()
+
+    scored_articles.sort(
+        key=lambda x: abs(_LABEL_TO_SCORE[x["label"]] * x["confidence"] * x["weight"]),
+        reverse=True,
+    )
+    top_headlines = " --- ".join(
+        f"[{a['label'].upper()} {a['confidence']:.0%}] {a['title']} ({a['provider']})"
+        for a in scored_articles[:5]
+    )
+
+    return (
+        f"Sentiment analysis for {comp_name} ({ticker.upper()}) across {len(articles)} recent articles: "
+        f"Composite score: {composite:+.4f} ({label}). "
+        f"Top influencing headlines: {top_headlines}"
+    )
+
+@tool("respond_to_greeting")
+def respond_to_greeting() -> str:
+    return "Hello! I'm a financial data agent. How can I assist you today?"
+
+@tool("respond_no_available_tool")
+def respond_no_available_tool(tool_name: str) -> str:
+    return f"Sorry, currently i'm capable of doing that. Check the list of avaiable tools for more information."