> ## Documentation Index
> Fetch the complete documentation index at: https://docs.llmrouter.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Tools Optimization

> Automatically filter out irrelevant tools and functions from your prompts to save tokens and prevent AI hallucinations.

***

When building complex AI agents, it's common to provide the model with a massive list of available functions (e.g., `search_database`, `send_email`, `calculate_math`, `create_ticket`).

However, sending 50 tool definitions in a single request when the user just said *"Hello"* causes two major problems:

1. **Token Waste:** Defining tools consumes a massive amount of input tokens.
2. **Hallucination Risk:** The more tools a model sees, the more likely it is to get confused and try to call the wrong tool.

LLM Router’s **Tools Optimization** engine solves this by actively scoring and filtering your `tools` array *before* sending it to the upstream model.

## How Tool Optimization Works

When a request containing a `tools` array arrives, our internal Gateway AI acts as a **Tool Selector Agent**.

1. It analyzes the user's prompt against the descriptions of all provided tools.
2. It assigns a **Relevance Score (0.0 to 1.0)** to each tool.
3. It detects **Dependencies** (e.g., if `create_element` requires `get_context`, both are scored highly).
4. If a tool's score falls below your configured threshold (`acceptScore`), it is **stripped from the request entirely**.

***

## Configuration

You configure this behavior inside the `gateway.toolOptimization` object.

```typescript TypeScript theme={null}
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.llmrouter.app/v1",
  apiKey: process.env.LLM_ROUTER_API_KEY,
});

async function main() {
  const response = await client.chat.completions.create({
    model: "claude-3-5-sonnet",
    messages: [{ role: "user", content: "What is 256 multiplied by 14?" }],

    // Imagine 20 different tools defined here
    tools: [
      {
        type: "function",
        function: { name: "calculator", description: "Performs math" },
      },
      {
        type: "function",
        function: { name: "send_email", description: "Sends an email" },
      },
      {
        type: "function",
        function: { name: "search_db", description: "Searches users" },
      },
    ],

    // @ts-expect-error - Custom LLM Router extension
    gateway: {
      toolOptimization: {
        enabled: true, // Master switch
        acceptScore: 0.5, // Only keep tools scoring 0.5 or higher
        alwaysInclude: ["search_db"], // These tools bypass the filter and are ALWAYS sent
      },
    },
  });

  console.log(response.choices[0].message.content);
}
main();
```

### What happens in this example?

1. The user asks a math question.
2. The Gateway scores `calculator` at **1.0**. It scores `send_email` at **0.0**.
3. It sees `search_db` in your `alwaysInclude` array.
4. **Result:** The router strips `send_email` and sends ONLY `calculator` and `search_db` to Claude. You save tokens, and Claude doesn't get distracted.

***

## Configuration Properties

### The `toolOptimization` Object

| Property        | Type       | Default | Description                                                                                                                                        |
| :-------------- | :--------- | :------ | :------------------------------------------------------------------------------------------------------------------------------------------------- |
| `enabled`       | `boolean`  | `true`  | Toggles the probabilistic tool filtering engine on or off.                                                                                         |
| `acceptScore`   | `number`   | `0.5`   | The minimum relevance score (0.0 to 1.0) required for a tool to be included in the final request to the LLM.                                       |
| `alwaysInclude` | `string[]` | `[]`    | An array of exact function names (e.g., `["get_weather"]`). These tools will **always** be sent to the LLM, bypassing the scoring engine entirely. |