# ===================================================================
# MODULE 6, LESSON 2: GENERATIVE AI AND LLMS FOR FINANCE
# ===================================================================

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import json
import re
from typing import List, Dict
import warnings
warnings.filterwarnings('ignore')

# Set style
sns.set_style("whitegrid")

print("="*70)
print("GENERATIVE AI AND LARGE LANGUAGE MODELS (LLMS) FOR FINANCE")
print("="*70)

# ----------------------------------------------------------------
# PART A: SIMULATED LLM RESPONSES (NO API KEY REQUIRED)
# ----------------------------------------------------------------

print("\n" + "-"*60)
print("PART A: Simulated LLM Responses")
print("-"*60)

print("""
Note: To use actual LLM APIs (OpenAI, Anthropic, etc.), you need an API key.
This demonstration uses simulated responses to illustrate concepts.

To run with a real API:
  1. Install: pip install openai
  2. Set your API key: openai.api_key = "your-key"
  3. Replace the simulate_llm_response function with actual API calls.
""")

def simulate_llm_response(prompt, model="simulated", temperature=0.7):
    """
    Simulate an LLM response (for demonstration).
    In practice, replace this with an actual API call.
    """
    # Keywords for demonstration
    if "summarise" in prompt.lower() or "summarize" in prompt.lower():
        return """
Summary:
The Federal Reserve's recent rate hike signals confidence in economic growth, 
though it has created volatility in equity markets. Technology stocks are 
particularly sensitive, while financials benefit from higher margins. 
The yield curve inversion suggests caution, but consumer spending remains resilient.
"""
    
    elif "sentiment" in prompt.lower():
        return """
Sentiment Analysis:
Overall sentiment: Positive (0.72)
- Positive indicators: record profits, strong demand, beating expectations
- Negative indicators: inflation fears, layoffs, recession warnings
- Neutral indicators: market volatility, interest rate changes
Confidence: 85%
"""
    
    elif "extract" in prompt.lower() or "json" in prompt.lower():
        return """
{
  "company": "Bank of America",
  "quarter": "Q4 2024",
  "revenue": "$25.3B",
  "profit": "$7.8B",
  "growth_rate": "12%",
  "key_drivers": ["investment banking", "wealth management"],
  "risks": ["inflation", "regulatory scrutiny"]
}
"""
    
    elif "think step-by-step" in prompt.lower() or "reasoning" in prompt.lower():
        return """
Step-by-Step Reasoning:
1. The company reported strong revenue growth (12% YoY) driven by investment banking.
2. Operating margins improved due to cost-cutting measures.
3. The balance sheet strengthened with increased capital reserves.
4. However, rising inflation poses a risk to consumer spending.
5. Recommendation: BUY with a target price of $45, based on a P/E of 15x.

Confidence: 78%
"""
    
    else:
        return "The Fed's actions have restored confidence in the banking sector, though challenges remain."

# ----------------------------------------------------------------
# PART B: PROMPT ENGINEERING EXAMPLES
# ----------------------------------------------------------------

print("\n" + "-"*60)
print("PART B: Prompt Engineering for Financial Tasks")
print("-"*60)

# Define sample prompts
prompts = {
    "Summarisation": "Summarise this earnings call transcript in 50 words: 'The company reported strong Q4 results with revenue of $25 billion, beating expectations. The CEO highlighted growth in digital banking and expansion into new markets. However, the company faces headwinds from rising interest rates and increased competition.'",
    
    "Sentiment Analysis": "Analyse the sentiment of this news headline: 'Fed rate hike sparks market rally as investors gain confidence.' Provide a score from -1 (negative) to +1 (positive).",
    
    "Extraction": "Extract the key financial metrics from this text and return as JSON: 'JPMorgan reported quarterly profit of $13.1 billion, up 8% from last year. Revenue reached $42.3 billion. The bank's CET1 ratio stood at 13.8%.'",
    
    "Chain-of-Thought": "Let's think step-by-step about whether to invest in Tesla. Consider: 1) Financial performance, 2) Market position, 3) Growth prospects, 4) Risks. Provide a recommendation.",
    
    "Role Prompting": "Act as a senior credit analyst. Assess the creditworthiness of a borrower with the following profile: revenue $5M, debt-to-income ratio 35%, credit score 620, industry: retail. Provide a recommendation."
}

# Simulate responses
for task, prompt in prompts.items():
    print(f"\n--- {task} ---")
    print(f"Prompt: {prompt[:100]}...")
    response = simulate_llm_response(prompt)
    print(f"Response: {response}")

# ----------------------------------------------------------------
# PART C: RETRIEVAL-AUGMENTED GENERATION (RAG) – CONCEPTUAL
# ----------------------------------------------------------------

print("\n" + "-"*60)
print("PART C: Retrieval-Augmented Generation (RAG)")
print("-"*60)

# Sample financial documents (simulated)
documents = [
    {"id": 1, "text": "The Federal Reserve raised interest rates to 5.25%, the highest level in 22 years."},
    {"id": 2, "text": "Bank of America reported Q3 earnings of $8.2 billion, beating analyst estimates."},
    {"id": 3, "text": "Inflation eased to 3.2% in October, down from 3.7% in September."},
    {"id": 4, "text": "The yield curve inverted, with the 10-year yield falling below the 2-year yield."},
    {"id": 5, "text": "Goldman Sachs upgraded Apple to 'buy', with a target price of $250."}
]

# Simulate embedding and retrieval (simple keyword matching)
def simple_retrieve(query, documents):
    """Simple retrieval based on keyword matching."""
    query_terms = query.lower().split()
    scored_docs = []
    for doc in documents:
        score = sum(1 for term in query_terms if term in doc['text'].lower())
        scored_docs.append((doc, score))
    # Sort by score (descending)
    scored_docs.sort(key=lambda x: x[1], reverse=True)
    return [doc[0] for doc in scored_docs if doc[1] > 0][:3]  # Top 3

# Sample query
query = "What is the current interest rate and how does it affect banks?"

# Retrieve relevant documents
retrieved_docs = simple_retrieve(query, documents)

print(f"Query: {query}")
print(f"\nRetrieved Documents ({len(retrieved_docs)}):")
for doc in retrieved_docs:
    print(f"  - {doc['text']}")

# Construct RAG prompt
rag_prompt = f"""
Context (from retrieved documents):
{chr(10).join([doc['text'] for doc in retrieved_docs])}

Question: {query}

Answer the question using the context above. If the context doesn't contain enough information, say so.
"""

print("\nRAG Prompt:")
print(rag_prompt)

print("\nRAG Response (simulated):")
print("Based on the available information, the Federal Reserve raised interest rates to 5.25%. This impacts banks by increasing their net interest margins (NIM), as they can earn more on loans. However, it also increases the cost of deposits and may lead to higher defaults. The inverted yield curve (10-year below 2-year) suggests market concerns about future economic growth, which could pressure bank profitability.")

# ----------------------------------------------------------------
# PART D: RISKS AND LIMITATIONS OF LLMS IN FINANCE
# ----------------------------------------------------------------

print("\n" + "-"*60)
print("PART D: Risks and Limitations of LLMs in Finance")
print("-"*60)

risks = [
    {
        "risk": "Hallucinations",
        "description": "LLMs can generate plausible but incorrect information.",
        "mitigation": "Use RAG with grounding in source documents; implement fact-checking; human-in-the-loop."
    },
    {
        "risk": "Bias and Fairness",
        "description": "Models can amplify biases present in training data.",
        "mitigation": "Audit outputs for bias; use diverse training data; implement bias detection."
    },
    {
        "risk": "Data Privacy",
        "description": "Sensitive financial data may be exposed to model providers.",
        "mitigation": "Use on-premise models (open-source); anonymise data; implement data governance."
    },
    {
        "risk": "Lack of Explainability",
        "description": "LLMs are black boxes; it's difficult to understand why a decision was made.",
        "mitigation": "Use model interpretability tools; document decision-making; regulatory oversight."
    },
    {
        "risk": "Regulatory Compliance",
        "description": "Regulators may not accept AI-generated outputs without human oversight.",
        "mitigation": "Maintain audit trails; human review for critical decisions; regulatory engagement."
    },
    {
        "risk": "Jailbreaking",
        "description": "Malicious prompts can bypass safety measures.",
        "mitigation": "Implement prompt filtering; use content moderation; monitor for misuse."
    }
]

risk_df = pd.DataFrame(risks)
print(risk_df.to_string(index=False))

print("\nBest Practices for LLM Adoption in Banking:")
print("""
1. Start with small-scale pilots in non-critical areas.
2. Implement robust testing and validation frameworks.
3. Maintain human oversight for all LLM-generated outputs.
4. Document all prompts, inputs, and outputs for audit purposes.
5. Use RAG to ground responses in verified data sources.
6. Regularly update models and monitor for drift.
7. Engage with regulators early in the adoption process.
8. Develop internal guidelines for responsible LLM use.
""")

# ----------------------------------------------------------------
# PART E: OPEN-SOURCE LLMS FOR FINANCE
# ----------------------------------------------------------------

print("\n" + "-"*60)
print("PART E: Open-Source LLMs for Finance")
print("-"*60)

open_source_models = [
    {
        "name": "Llama 2 / Llama 3",
        "developer": "Meta",
        "size": "7B-70B",
        "advantages": "Open weights; strong performance; customisable.",
        "use_case": "General-purpose financial analysis, document processing."
    },
    {
        "name": "Mistral 7B",
        "developer": "Mistral AI",
        "size": "7B",
        "advantages": "Efficient; high performance; Apache 2.0 license.",
        "use_case": "Cost-effective deployment; high accuracy for size."
    },
    {
        "name": "Falcon",
        "developer": "Technology Innovation Institute",
        "size": "7B-180B",
        "advantages": "Open weights; good for Arabic and multilingual.",
        "use_case": "Financial analytics in the Middle East and Africa."
    },
    {
        "name": "BloombergGPT",
        "developer": "Bloomberg",
        "size": "50B",
        "advantages": "Trained on 363 billion tokens of financial data; domain-specific.",
        "use_case": "Financial sentiment, news analysis, regulatory filings."
    },
    {
        "name": "FinGPT",
        "developer": "AI4Finance",
        "size": "Various",
        "advantages": "Open-source; specialised for finance; constantly updated.",
        "use_case": "Financial news summarisation, sentiment analysis, data extraction."
    }
]

model_df = pd.DataFrame(open_source_models)
print(model_df.to_string(index=False))

# ----------------------------------------------------------------
# PART F: BUSINESS APPLICATIONS AND IMPLEMENTATION ROADMAP
# ----------------------------------------------------------------

print("\n" + "="*70)
print("PART F: Business Applications and Implementation Roadmap")
print("="*70)

applications = {
    "Customer Service Automation": {
        "Description": "AI chatbots for customer queries, account information, and transactions.",
        "Implementation": "Use GPT-4 or open-source model with RAG on product documents.",
        "Timeline": "3-6 months"
    },
    "Document Processing": {
        "Description": "Extract structured data from financial documents (invoices, contracts, filings).",
        "Implementation": "Use GPT-4 with structured output prompting or fine-tuned open-source model.",
        "Timeline": "2-4 months"
    },
    "Research Assistance": {
        "Description": "Summarise earnings calls, analyst reports, and news for investment decisions.",
        "Implementation": "Use LLMs with RAG on internal research databases.",
        "Timeline": "1-3 months"
    },
    "Regulatory Compliance": {
        "Description": "Check documents for compliance, identify regulatory risks.",
        "Implementation": "Use LLMs with fine-tuning on regulatory text and RAG on regulatory guidelines.",
        "Timeline": "6-12 months"
    },
    "Trading Signals": {
        "Description": "Generate trading signals from news, social media, and alternative data.",
        "Implementation": "Use sentiment analysis (FinBERT) + generative AI for summarisation.",
        "Timeline": "6-12 months"
    },
    "Code Generation": {
        "Description": "Assist data scientists and developers with code for financial analysis.",
        "Implementation": "Use GitHub Copilot, ChatGPT, or open-source code models.",
        "Timeline": "1-2 months"
    }
}

app_df = pd.DataFrame([
    {
        "Application": app,
        "Description": details["Description"],
        "Implementation": details["Implementation"],
        "Timeline": details["Timeline"]
    }
    for app, details in applications.items()
])
print(app_df.to_string(index=False))

print("\nImplementation Roadmap:")
print("""
Phase 1 (Months 1-3): Pilot Projects
  - Select 1-2 low-risk use cases (e.g., summarisation, code assistance).
  - Use API-based models (OpenAI, Anthropic) for speed.
  - Establish evaluation metrics and human oversight.

Phase 2 (Months 4-6): Scale and Integrate
  - Expand to more use cases.
  - Implement RAG for proprietary data.
  - Begin evaluating open-source models for cost savings.

Phase 3 (Months 7-12): Production and Governance
  - Deploy models in production with monitoring and auditing.
  - Establish internal policies for LLM use.
  - Engage with regulators for compliance.

Phase 4 (Months 12+): Advanced and Proprietary
  - Fine-tune open-source models on proprietary data.
  - Develop in-house LLM capabilities (if justified).
  - Explore multi-modal capabilities (text + numbers + charts).
""")