GenAI Hands-On Workshops

Goal: Build an agent that can use tools to perform specific tasks.

🔧 Tool Definition

from langchain_core.tools import tool

@tool
def calculate_square(number: str) -> str:
    """Calculates the square of an integer.
    Use when user asks to calculate square of a number."""
    number = int(number.strip())
    return f"The square of {number} is {number ** 2}."

⚡ Agent Decision Flow

📤 Expected Output

User: What is the square of 7?
→ Agent decides to use calculate_square tool
→ Tool returns: "The square of 7 is 49."
Assistant: The square of 7 is 49.

python3 02_agent_with_tools.py

Goal: Create a searchable knowledge base using FAISS.

📊 RAG Pipeline

💻 Code Example

from sentence_transformers import SentenceTransformer
import faiss

# Load embedding model
model = SentenceTransformer('all-MiniLM-L6-v2')

# Generate embeddings
texts = ["Critical TechWorks Principles...", "Quality and collaboration..."]
embeddings = model.encode(texts)

# Create FAISS index
index = faiss.IndexFlatIP(384)  # 384 = embedding dimension
index.add(embeddings)
faiss.write_index(index, "faiss.index")

📁 Generated Files

faiss_index/
├── faiss.index      # Vector index (binary)
├── embeddings.npy   # Embedding vectors
└── texts.json       # Original text chunks

python3 03_rag_create_vector_database.py

Goal: Query the vector database with natural language.

💻 Code Example

# Encode user query
query = "What are Critical TechWorks principles?"
q_embedding = model.encode([query])

# Search in FAISS
scores, indices = index.search(q_embedding, k=3)  # Top 3 results

# Display results
for score, idx in zip(scores[0], indices[0]):
    print(f"Score: {score:.3f} → {texts[idx][:100]}...")

📤 Expected Output

Query: "principles"

Result 1 (Score: 0.85): Critical TechWorks Principles: A principle is a rule...
Result 2 (Score: 0.72): Quality and collaboration are fundamental...
Result 3 (Score: 0.68): Our core values include transparency...

🧪 Try It Yourself

1. Search for "principles" - should find relevant docs
2. Search for "CI/CD" - should find quality docs
3. Search for "pizza" - should return low scores (irrelevant)

python3 04_rag_semantic_search.py

Goal: Build a production-ready chatbot with LangGraph StateGraph and output guardrails.

📊 Architecture (8 Nodes)

💻 Building the Graph

from langgraph.graph import StateGraph, END

graph = StateGraph(AgentState)

# Add nodes
graph.add_node("classify", classify_intent)
graph.add_node("calculate", use_calculate_square)
graph.add_node("search", use_faiss_search)
graph.add_node("guardrails", check_guardrails)

# Set entry point and routing
graph.set_entry_point("classify")
graph.add_conditional_edges("classify", route_by_intent, {...})

# Compile
agent = graph.compile()

🛡️ Guardrails Implementation

🚫 Blocked Response Example

User: "How to hack a computer?"

🔍 [CLASSIFY] → chat
💬 [GENERATE] → Draft generated
🛡️ [GUARDRAILS] → ❌ BLOCKED
🚫 Response: "I cannot provide information on that topic..."

🧪 Try It Yourself

• "What is the square of 8?" → ✅ PASSES guardrails
• "Tell me about Critical TechWorks" → ✅ PASSES guardrails
• "How to hack a computer?" → 🚫 BLOCKED by guardrails
• "How to steal passwords?" → 🚫 BLOCKED by guardrails

python3 05_complete_agent_with_rag.py

def check_answer_type(state: State) -> dict:
    question = state["messages"][-1].content
    needs_math = "YES" == llm.invoke(prompt).upper().strip()
    return {"needs_math": needs_math}

def math_answer(state: State) -> dict:
    result = math_tools[operation](*numbers)
    return {"messages": [AIMessage(content=f"Result: {result}")]}

def route(state: State) -> str:
    if state["needs_math"]:
        return "math_answer"
    return "general_answer"

graph.add_conditional_edges(
    "check_answer_type",
    route,
    {"math_answer": "math_answer", "general_answer": "general_answer"}
)

from langgraph.graph import StateGraph, END

graph = StateGraph(State)
graph.add_node("check_answer_type", check_answer_type)
graph.add_node("math_answer", math_answer)
graph.add_node("general_answer", general_answer)

graph.set_entry_point("check_answer_type")
graph.add_edge("math_answer", END)
graph.add_edge("general_answer", END)

app = graph.compile()

from langchain_mcp_adapters.client import MultiServerMCPClient

client = MultiServerMCPClient({
    "celonis_mcp": {
        "transport": "http",
        "url": "http://localhost:8000/mcp",
    },
    "jira_mcp": {
        "transport": "stdio",
        "command": sys.executable,
        "args": ["mcp/server.py", "stdio"],
    }
})

mcp_tools = await client.get_tools()

def supervisor(state: GraphState) -> dict:
    # Analyze chat history and execution log
    # Decide: "answer_generation" or "tool_execution"
    response = json.loads(llm.invoke(prompt))
    return {"supervisor_decision": response["supervisor_decision"]}