Context-Aware Tools

Most tools are "Static"—they take an input and return an output regardless of who is asking. However, a production agent needs Context-Aware Tools.

• A search_docs tool should know if the user is a Developer or a Manager.
• A get_balance tool should automatically know which account_id belongs to the current thread.

In this lesson, we will learn how to "Secretly" pass state into our tools so the agent doesn't have to worry about the plumbing.

---

1. The Hidden Parameter Pattern

When an LLM calls a tool, it only provides the parameters defined in the schema (e.g., query). It doesn't know about user_id or session_token.

We use Injected Parameters to bridge this gap.

The Implementation

1. The Orchestrator (LangGraph) sees a tool call.
2. Before executing the tool, it "Enriches" the arguments with data from the State.

# The LLM thinks the tool only takes 'query'
# But the Python function also takes 'authenticated_user_id'
def search_private_docs(query: str, state_user_id: str):
    # This logic ensures the agent can't search other users' files
    return db.search(query, owner=state_user_id)

---

2. Tools that "Read" the Chat History

Sometimes, a tool needs to know what was said 5 minutes ago to be accurate.

Example: The "Clarification" Tool

Say an agent has a tool execute_sql. If the tool sees that the user previously said "Show me ONLY North American sales," the tool can automatically add WHERE region = 'NA' to every query the agent tries to run, even if the agent forgot to include it.

---

3. The "Dynamic Schema" Strategy

You can change what tools the agent sees based on the state.

State Condition	Tool List
User is Unauthenticated	login, search_public_faqs
User is Admin	get_user_logs, delete_database, reset_password
Task is in "Drafting" mode	write_snippet, search_vocabulary

Why? By limiting the toolset dynamically, you reduce the model's confusion and prevent unauthorized tool access.

---

4. Context-Aware Error Messages

A context-aware tool doesn't just return a generic error. It looks at the goal.

• Generic: Error: Permission Denied.
• Context-Aware: Error: You are trying to access the 'CEO_Bonuses' table. Your current user (John) only has 'Staff' level access. Please try searching for 'Standard_Salaries' instead.

---

5. Implementation: LangChain RunnableConfig

In LangChain, we use the config object to pass context through the system without the LLM seeing it.

from langchain_core.runnables import RunnableConfig

async def my_tool(query: str, config: RunnableConfig):
    # Retrieve data injected by the API layer (Module 10.3)
    user_id = config.get("configurable", {}).get("user_id")
    return await db.lookup(query, user_id)

---

6. The "Environment Aware" Tool

A tool can adjust its behavior based on the Time or Location.

• search_news tool → automatically appends published_after: [Today's Date].
• book_flight tool → automatically sets the origin_city to the user's current GPS location from the state.

---

Summary and Mental Model

Think of a Context-Aware Tool like a Smart Assistant.

• A Dumb Assistant asks: "Who should I call?" (Requires explicit input).
• A Smart Assistant sees you are in a meeting with John and says: "I see you're with John, should I call his office?" (Uses context).

The more context your tools have, the less work the LLM has to do.

---

Exercise: Context Injection

1. The Scenario: You are building an agent for a Law Firm.
   • The search_contracts tool must only return documents for the specific client_id stored in the LangGraph state.
   • Write the Python function signature for this tool, identifying which parameter is for the LLM and which is for the State.
2. Security: Why is it safer to inject the user_id from the Config rather than having the agent provide it as a schema parameter?
   • (Hint: Could an agent hallucinate the wrong user_id?)
3. Logic: If an agent is in "Help Mode," how would you change the Search tool to return "Simple Explanations" instead of "Raw Code"? Ready to manage hundreds of tools? Let's move to High-Cardinality Tool Sets.