Agent Tools

This app uses a LangGraph ReAct agent with LangChain tools (defined in backend/chatbot_agent.py). Tools let the model call out to functions like search, APIs, or utilities.

What you get out-of-the-box

web_search(query): fetches web results via Tavily (requires TAVILY_API_KEY).

Create the agent:

from chatbot_agent import create_chatbot_agent
agent = create_chatbot_agent(model_name="mistralai/Mistral-Small-24B-Instruct")

Current Implementation

The ReAct agent currently has one tool available:

@tool
def web_search(query: str) -> str:
    """
    Do a web search to query additional information for the user.
    """
    logger.info(f"Searching web for: {query}")
    retriever = TavilySearchAPIRetriever(k=8)
    chunks = retriever.invoke(query)
    logger.info(chunks)
    return "\n\n".join(chunk.page_content for chunk in chunks)

The agent is created with:

return create_react_agent(model=llm, prompt=_prompt, tools=[web_search], state_schema=AgentStatePydantic)

Add a new tool (step‑by‑step)

Define a function with a helpful docstring and decorate with @tool.
Keep inputs simple (prefer a single query: str).
Handle timeouts and exceptions; return a readable string on failure.
Register the tool in create_chatbot_agent by adding it to the tools list.

Example – a simple clock tool:

from langchain_core.tools import tool
 
@tool
def get_current_time(_: str) -> str:
    """Return the current UTC time in ISO 8601 format."""
    import datetime
    return datetime.datetime.utcnow().isoformat()
 
# In create_chatbot_agent(...), include `get_current_time` in the tools list.
# Example: tools=[web_search, get_current_time] in backend/chatbot_agent.py

Best practices

Keep tool outputs short and factual; avoid speculation.
Validate and sanitize inputs (URLs, IDs).
Timeouts and retries: be defensive against slow or flaky APIs.
Don’t log secrets; read them from env (e.g., IONOS_API_KEY, TAVILY_API_KEY).
Cache where it helps (e.g., web results) to reduce latency and cost.

Setup API