LangChain Tools — Preventing Hallucinated Tool Loops

from langchain_core.tools import tool
from langchain_openai import ChatOpenAI
from pydantic import BaseModel, Field
import json

# --- Step 1: Define a strict input schema with Pydantic ---
# This schema becomes JSON Schema that gets sent to the LLM.
# Field descriptions are part of what the LLM reads to understand how to call this.
class StockLookupInput(BaseModel):
    ticker: str = Field(
        description="The stock ticker symbol in uppercase, e.g. 'AAPL' or 'MSFT'"
    )
    metric: str = Field(
        description="The metric to retrieve. One of: 'price', 'pe_ratio', 'market_cap'"
    )

# --- Step 2: Decorate with @tool and provide a rich description ---
# The description is the LLM's only guide for WHEN to use this.
# Be specific: tell it what it returns and when NOT to use it.
@tool(args_schema=StockLookupInput)
def get_stock_metric(ticker: str, metric: str) -> str:
    """
    Retrieve a real-time financial metric for a publicly traded stock.
    Use this when the user asks about current stock prices, P/E ratios,
    or market capitalisation. Do NOT use this for historical data or crypto.
    Returns a formatted string with the value and currency where applicable.
    """
    # Simulated data store — in production this would call a financial API
    mock_data = {
        "AAPL": {"price": "$189.43", "pe_ratio": "31.2", "market_cap": "$2.94T"},
        "MSFT": {"price": "$415.61", "pe_ratio": "36.8", "market_cap": "$3.08T"},
    }
    ticker = ticker.upper()
    if ticker not in mock_data:
        # Return a clear error string — never raise inside a tool unless you want
        # the agent to crash. Return errors as strings so the LLM can reason about them.
        return f"Ticker '{ticker}' not found. Available tickers: AAPL, MSFT"
    if metric not in mock_data[ticker]:
        return f"Metric '{metric}' is invalid. Valid options: price, pe_ratio, market_cap"
    return f"{ticker} {metric}: {mock_data[ticker][metric]}"

# --- Step 3: Inspect what the LLM actually sees ---
# This is the JSON Schema injected into the model's context.
print("=== Tool Name ===")
print(get_stock_metric.name)  # 'get_stock_metric'

print("\n=== Tool Description (what LLM reads) ===")
print(get_stock_metric.description)

print("\n=== Args Schema (sent as JSON Schema to the model) ===")
print(json.dumps(get_stock_metric.args_schema.model_json_schema(), indent=2))

# --- Step 4: Bind tool to model and observe the raw tool_call output ---
llm = ChatOpenAI(model="gpt-4o", temperature=0)
llm_with_tools = llm.bind_tools([get_stock_metric])

response = llm_with_tools.invoke("What's Apple's current P/E ratio?")

print("\n=== AIMessage tool_calls (raw LLM output — no execution yet) ===")
for tc in response.tool_calls:
    print(json.dumps(tc, indent=2))

# --- Step 5: Manually execute the tool call ---
# In production an agent loop or LangGraph node does this automatically.
tool_call = response.tool_calls[0]
result = get_stock_metric.invoke(tool_call["args"])
print("\n=== Tool Execution Result ===")
print(result)

from langchain_core.tools import BaseTool
from pydantic import BaseModel, Field, field_validator
from typing import Optional, Type
import httpx
import hashlib
import time

# --- Input schema with built-in validation ---
class WeatherQueryInput(BaseModel):
    city: str = Field(description="City name to get weather for, e.g. 'London' or 'Tokyo'")
    units: str = Field(
        default="celsius",
        description="Temperature units: 'celsius' or 'fahrenheit'"
    )

    # Pydantic v2 validator — catches bad input BEFORE the tool runs
    @field_validator("units")
    @classmethod
    def validate_units(cls, value: str) -> str:
        allowed = {"celsius", "fahrenheit"}
        if value.lower() not in allowed:
            raise ValueError(f"Units must be one of {allowed}, got '{value}'")
        return value.lower()

    @field_validator("city")
    @classmethod
    def validate_city_name(cls, value: str) -> str:
        # Prevent prompt injection via city name
        if len(value) > 100 or not value.replace(" ", "").replace("-", "").isalpha():
            raise ValueError("City name contains invalid characters")
        return value.strip().title()

# --- Production BaseTool with dependency injection ---
class WeatherTool(BaseTool):
    name: str = "get_current_weather"
    description: str = (
        "Retrieve current weather conditions for a city. "
        "Use this when the user asks about today's weather, temperature, "
        "or current conditions in a specific city. "
        "Returns temperature and a brief condition description. "
        "Do NOT use for forecasts — only current conditions."
    )
    args_schema: Type[BaseModel] = WeatherQueryInput

    # Injected dependencies — passed at construction, not module-scope globals
    api_key: str
    base_url: str = "https://api.openweathermap.org/data/2.5"
    request_timeout: int = 5

    # Simple in-memory deduplication to prevent redundant API calls
    # In production, use Redis with TTL instead
    _call_cache: dict = {}

    def _make_cache_key(self, city: str, units: str) -> str:
        """Cache key includes a time bucket so we don't serve stale data"""
        # Round to nearest 10-minute window for caching
        time_bucket = int(time.time() / 600)
        raw = f"{city}:{units}:{time_bucket}"
        return hashlib.md5(raw.encode()).hexdigest()

    def _run(self, city: str, units: str = "celsius") -> str:
        """
        Synchronous execution path. LangChain calls this when the tool
        is invoked in a non-async context.
        """
        cache_key = self._make_cache_key(city, units)
        if cache_key in self._call_cache:
            print(f"[WeatherTool] Cache HIT for {city}")
            return self._call_cache[cache_key]

        # Map our friendly unit names to the API's expected values
        api_units = "metric" if units == "celsius" else "imperial"
        unit_symbol = "°C" if units == "celsius" else "°F"

        try:
            response = httpx.get(
                f"{self.base_url}/weather",
                params={"q": city, "appid": self.api_key, "units": api_units},
                timeout=self.request_timeout,
            )
            # Don't raise on 404 — return a useful string so the LLM can handle it
            if response.status_code == 404:
                return f"City '{city}' not found. Check the spelling and try again."

            response.raise_for_status()  # Raise on 5xx errors
            data = response.json()

            temp = data["main"]["temp"]
            condition = data["weather"][0]["description"]
            humidity = data["main"]["humidity"]
            result = f"{city}: {temp}{unit_symbol}, {condition}, humidity {humidity}%"

            # Cache the successful result
            self._call_cache[cache_key] = result
            return result

        except httpx.TimeoutException:
            # Return recoverable error — LLM can retry or inform user
            return f"Weather service timed out after {self.request_timeout}s. Try again shortly."
        except httpx.HTTPStatusError as exc:
            return f"Weather API error (HTTP {exc.response.status_code}). Service may be down."

    async def _arun(self, city: str, units: str = "celsius") -> str:
        """
        Async execution path — called when the agent runs in an async context.
        Always implement both _run and _arun in production tools.
        Using httpx.AsyncClient here for true non-blocking IO.
        """
        cache_key = self._make_cache_key(city, units)
        if cache_key in self._call_cache:
            return self._call_cache[cache_key]

        api_units = "metric" if units == "celsius" else "imperial"
        unit_symbol = "°C" if units == "celsius" else "°F"

        async with httpx.AsyncClient(timeout=self.request_timeout) as client:
            try:
                response = await client.get(
                    f"{self.base_url}/weather",
                    params={"q": city, "appid": self.api_key, "units": api_units},
                )
                if response.status_code == 404:
                    return f"City '{city}' not found."
                response.raise_for_status()
                data = response.json()
                temp = data["main"]["temp"]
                condition = data["weather"][0]["description"]
                humidity = data["main"]["humidity"]
                result = f"{city}: {temp}{unit_symbol}, {condition}, humidity {humidity}%"
                self._call_cache[cache_key] = result
                return result
            except httpx.TimeoutException:
                return "Weather service timed out. Please retry."

# --- Constructing the tool with injected config ---
# In a real app, api_key comes from environment variables or a secrets manager
weather_tool = WeatherTool(
    api_key="your-openweathermap-api-key",
    request_timeout=8
)

# --- Direct invocation test (no LLM needed) ---
result = weather_tool.invoke({"city": "London", "units": "celsius"})
print(result)

# --- Test validation catches bad input cleanly ---
try:
    weather_tool.invoke({"city": "London", "units": "kelvin"})
except Exception as e:
    print(f"Validation caught: {e}")

from typing import Annotated, Sequence
from langchain_core.messages import BaseMessage, HumanMessage, AIMessage
from langchain_core.tools import tool
from langchain_openai import ChatOpenAI
from langgraph.graph import StateGraph, END
from langgraph.graph.message import add_messages
from langgraph.prebuilt import ToolNode
from pydantic import BaseModel
import operator

# --- Define the tools our agent can use ---

@tool
def search_product_catalog(query: str) -> str:
    """
    Search the internal product catalog for items matching a description.
    Use this when the user asks what products we carry or wants to find
    a specific item. Returns product names, IDs, and prices.
    """
    # Simulated catalog — real version would query a vector store or DB
    catalog = [
        {"id": "P001", "name": "Wireless Keyboard", "price": 79.99},
        {"id": "P002", "name": "Ergonomic Mouse", "price": 49.99},
        {"id": "P003", "name": "USB-C Hub", "price": 34.99},
    ]
    query_lower = query.lower()
    matches = [
        f"{p['name']} (ID: {p['id']}, ${p['price']})"
        for p in catalog
        if any(word in p["name"].lower() for word in query_lower.split())
    ]
    if not matches:
        return f"No products found matching '{query}'. Try broader search terms."
    return "Found: " + ", ".join(matches)

@tool
def check_inventory(product_id: str) -> str:
    """
    Check real-time stock levels for a product by its ID.
    Use this AFTER searching the catalog to get a product ID.
    Returns stock count and estimated restock date if out of stock.
    """
    inventory = {
        "P001": {"stock": 143, "restock": None},
        "P002": {"stock": 0, "restock": "2025-09-15"},
        "P003": {"stock": 27, "restock": None},
    }
    product_id = product_id.upper()
    if product_id not in inventory:
        return f"Product ID '{product_id}' not recognised in inventory system."

    item = inventory[product_id]
    if item["stock"] == 0:
        return f"{product_id}: OUT OF STOCK. Expected restock: {item['restock']}"
    return f"{product_id}: {item['stock']} units in stock."

# Collect all tools — ToolNode needs this list to route calls by name
available_tools = [search_product_catalog, check_inventory]

# --- Define graph state ---
# Annotated with add_messages reducer: new messages are APPENDED, not overwritten
# This is critical — without the reducer you'd lose conversation history on each step
class AgentState(BaseModel):
    messages: Annotated[Sequence[BaseMessage], add_messages]
    iteration_count: int = 0  # Guard against infinite loops
    MAX_ITERATIONS: int = 10  # Class constant baked into state type

    class Config:
        arbitrary_types_allowed = True

# --- Wire up LLM with tools ---
catalog_llm = ChatOpenAI(model="gpt-4o", temperature=0)
catalog_llm_with_tools = catalog_llm.bind_tools(available_tools)

# --- Node 1: Model node — calls the LLM and returns its response ---
def call_model_node(state: AgentState) -> dict:
    """
    This node calls the LLM. It reads the full message history from state
    and returns the new AIMessage. LangGraph merges this with existing messages
    via the add_messages reducer.
    """
    current_iteration = state.iteration_count + 1
    print(f"[Agent] Model call — iteration {current_iteration}")

    # Safety valve: if we've looped too many times, force a stop
    if current_iteration >= state.MAX_ITERATIONS:
        from langchain_core.messages import AIMessage
        forced_stop = AIMessage(
            content="I've reached the maximum number of steps. "
                    "Here's what I found so far based on available information."
        )
        return {"messages": [forced_stop], "iteration_count": current_iteration}

    response = catalog_llm_with_tools.invoke(state.messages)
    return {"messages": [response], "iteration_count": current_iteration}

# --- Node 2: ToolNode — executes all tool_calls from the last AIMessage ---
# ToolNode handles parallel execution automatically when multiple tools are called
tool_executor_node = ToolNode(tools=available_tools)

# --- Routing function: decide whether to use a tool or end ---
def should_continue_routing(state: AgentState) -> str:
    """
    Conditional edge function. Inspects the last message:
    - If it has tool_calls -> route to tools node
    - Otherwise -> route to END (agent has its final answer)
    """
    last_message = state.messages[-1]
    # isinstance check ensures we're looking at an AIMessage, not a ToolMessage
    if isinstance(last_message, AIMessage) and last_message.tool_calls:
        print(f"[Router] Routing to tools: {[tc['name'] for tc in last_message.tool_calls]}")
        return "use_tools"
    print("[Router] No tool calls — agent finished")
    return "end"

# --- Build the graph ---
agent_graph = StateGraph(AgentState)

# Register nodes
agent_graph.add_node("model", call_model_node)
agent_graph.add_node("tools", tool_executor_node)

# Set entry point
agent_graph.set_entry_point("model")

# Conditional edge from model: either call tools or finish
agent_graph.add_conditional_edges(
    "model",
    should_continue_routing,
    {
        "use_tools": "tools",  # route name -> node name
        "end": END,
    },
)

# After tools execute, always go back to model to process results
agent_graph.add_edge("tools", "model")

# Compile into a runnable
runnable_agent = agent_graph.compile()

# --- Run the agent ---
print("=== STARTING AGENT ===")
initial_state = {"messages": [HumanMessage(content="Do you have any keyboards in stock?")]}

final_state = runnable_agent.invoke(initial_state)

print("\n=== FINAL ANSWER ===")
print(final_state["messages"][-1].content)
print(f"Total iterations: {final_state['iteration_count']}")

import asyncio
from langchain_core.tools import tool
from langchain_core.messages import HumanMessage
from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent
import time

# --- Two slow tools that should run in PARALLEL, not sequentially ---

@tool
async def fetch_user_profile(user_id: str) -> str:
    """
    Fetch a user's profile from the user service by their ID.
    Returns name, email, and account tier. Use when personalising responses.
    """
    await asyncio.sleep(0.8)  # Simulates 800ms database query
    profiles = {
        "user_42": {"name": "Sarah Chen", "email": "sarah@example.com", "tier": "premium"},
        "user_99": {"name": "James Okafor", "email": "james@example.com", "tier": "free"},
    }
    if user_id not in profiles:
        return f"User '{user_id}' not found"
    p = profiles[user_id]
    return f"Name: {p['name']}, Email: {p['email']}, Tier: {p['tier']}"

@tool
async def fetch_recent_orders(user_id: str) -> str:
    """
    Fetch the 3 most recent orders for a user.
    Returns order IDs, items, and total amounts.
    Use when user asks about their order history or past purchases.
    """
    await asyncio.sleep(1.0)  # Simulates 1s API call to order service
    orders = {
        "user_42": [
            {"id": "ORD-8821", "item": "Wireless Keyboard", "total": 79.99},
            {"id": "ORD-7703", "item": "USB-C Hub", "total": 34.99},
        ],
        "user_99": [{"id": "ORD-9011", "item": "Ergonomic Mouse", "total": 49.99}],
    }
    if user_id not in orders:
        return f"No orders found for user '{user_id}'"
    order_list = ", ".join(
        f"{o['id']}: {o['item']} (${o['total']})" for o in orders[user_id]
    )
    return f"Recent orders: {order_list}"

async def demonstrate_parallel_execution():
    """
    With parallel tool calling, both fetch_user_profile and fetch_recent_orders
    run concurrently. Total time should be ~1s (the slower tool), not ~1.8s (sum).
    """
    parallel_tools = [fetch_user_profile, fetch_recent_orders]
    agent_llm = ChatOpenAI(model="gpt-4o", temperature=0)

    # create_react_agent is a convenience wrapper that builds the LangGraph pattern
    # we built manually above — useful for standard use cases
    react_agent = create_react_agent(agent_llm, parallel_tools)

    user_query = (
        "Give me a personalised summary for user_42. "
        "I need their profile AND their recent orders."
    )

    start_time = time.perf_counter()
    result = await react_agent.ainvoke(
        {"messages": [HumanMessage(content=user_query)]}
    )
    elapsed = time.perf_counter() - start_time

    print(f"Execution time: {elapsed:.2f}s")
    print(f"(If tools ran sequentially this would be ~1.8s — parallel saves ~0.8s)")
    print("\nFinal response:")
    print(result["messages"][-1].content)

async def demonstrate_streaming_events():
    """
    Stream agent events to show the user what's happening in real time.
    This is what you'd use to build a live 'thinking...' UI indicator.
    """
    streaming_tools = [fetch_user_profile]
    stream_llm = ChatOpenAI(model="gpt-4o", temperature=0)
    streaming_agent = create_react_agent(stream_llm, streaming_tools)

    print("\n=== STREAMING EVENTS ===")
    async for event in streaming_agent.astream_events(
        {"messages": [HumanMessage(content="Look up profile for user_42")]},
        version="v2",  # v2 is the current stable events API
    ):
        event_kind = event["event"]

        if event_kind == "on_chat_model_stream":
            # Stream LLM token output as it arrives
            token = event["data"]["chunk"].content
            if token:  # Filter empty chunks
                print(token, end="", flush=True)

        elif event_kind == "on_tool_start":
            # Notify UI that a tool is being called
            print(f"\n[UI INDICATOR] Calling tool: {event['name']}...")

        elif event_kind == "on_tool_end":
            # Tool finished — you could update a progress indicator
            print(f"[UI INDICATOR] Tool '{event['name']}' completed")

async def main():
    await demonstrate_parallel_execution()
    await demonstrate_streaming_events()

asyncio.run(main())