import { Tool } from "../types.js";
/**
* Token thresholds for warnings
*/
export const TOKEN_THRESHOLDS = {
SMALL: 1000, // No warning needed
MEDIUM: 5000, // Mild warning
LARGE: 10000, // Strong warning
HUGE: 50000, // Critical warning - requires confirmation
MASSIVE: 100000, // Extreme warning - strongly discourage
} as const;
/**
* Get warning level based on token count
*/
export function getTokenWarningLevel(tokens: number): "none" | "mild" | "strong" | "critical" | "extreme" {
if (tokens < TOKEN_THRESHOLDS.SMALL) return "none";
if (tokens < TOKEN_THRESHOLDS.MEDIUM) return "mild";
if (tokens < TOKEN_THRESHOLDS.LARGE) return "strong";
if (tokens < TOKEN_THRESHOLDS.HUGE) return "critical";
return "extreme";
}
/**
* Generate warning message for token consumption
*/
export function generateTokenWarning(
tokens: number,
operation: string,
itemCount?: number
): string {
const level = getTokenWarningLevel(tokens);
if (level === "none") return "";
const itemsText = itemCount ? ` for ${itemCount} items` : "";
switch (level) {
case "mild":
return `\n📊 **Note:** This ${operation}${itemsText} will use ~${tokens.toLocaleString()} tokens.`;
case "strong":
return `\n⚠️ **Warning:** This ${operation}${itemsText} will consume ~${tokens.toLocaleString()} tokens (significant context usage).`;
case "critical":
return `\n🚨 **Critical Warning:** This ${operation}${itemsText} will consume ~${tokens.toLocaleString()} tokens!\n` +
`This is a very large amount that may use up most of your context window.\n` +
`Consider:\n` +
`• Limiting your search/request to fewer items\n` +
`• Using specific filters to reduce results\n` +
`• Requesting content without full text (summaries only)`;
case "extreme":
return `\n🛑 **EXTREME WARNING:** This ${operation}${itemsText} would consume ~${tokens.toLocaleString()} tokens!\n` +
`This exceeds typical LLM context windows and is strongly discouraged.\n` +
`**Recommendations:**\n` +
`• Break down your request into smaller chunks\n` +
`• Use more specific search queries\n` +
`• Request summaries instead of full content\n` +
`• Use the batch_get_content tool with specific IDs`;
}
}
/**
* Generate a preview response when content is too large
*/
export function generatePreviewResponse(
content: any,
actualTokens: number,
operation: string
): string {
let response = `# ⚠️ Large Content Warning\n\n`;
response += `The ${operation} would return ~${actualTokens.toLocaleString()} tokens.\n\n`;
response += `## Options:\n`;
response += `1. **Get summary only** - Use without \`includeFullContent\` flag\n`;
response += `2. **Get specific sections** - Use the \`smart_sections\` tool\n`;
response += `3. **Accept full content** - Re-run with \`acceptLargeContent: true\` parameter\n\n`;
response += `## Content Preview:\n`;
response += `**Title:** ${content.title || 'Untitled'}\n`;
response += `**Type:** ${content.contentType}\n`;
response += `**URL:** ${content.url}\n`;
if (content.summary) {
const summaryObj = typeof content.summary === 'string'
? { text: content.summary }
: content.summary;
if (summaryObj.one_sentence) {
response += `\n**Summary:** ${summaryObj.one_sentence}\n`;
}
}
return response;
}
/**
* Instructions for LLMs on how to handle large content requests
*/
export const CONTEXT_MANAGEMENT_INSTRUCTIONS = `
## Context Management Instructions for LLMs
When users request content from Noverload, be mindful of token consumption:
### Before Returning Large Content:
1. **Check Token Count**: If content exceeds 10,000 tokens, warn the user
2. **Suggest Alternatives**:
- Use summaries instead of full text
- Use smart_sections for specific parts
- Filter searches more narrowly
3. **Require Confirmation**: For content over 50,000 tokens, ask for explicit confirmation
### Best Practices:
- Start with summaries and metadata
- Only fetch full content when specifically needed
- Use batch operations efficiently
- Guide users to more targeted queries
### Warning Thresholds:
- < 1,000 tokens: No warning needed
- 1,000-5,000 tokens: Note the size
- 5,000-10,000 tokens: Mild warning
- 10,000-50,000 tokens: Strong warning with alternatives
- > 50,000 tokens: Require explicit confirmation
### Example User Interaction:
User: "Show me all my saved content about AI"
Assistant: "I found 45 items about AI. Getting full content for all would use ~250,000 tokens.
Would you like to:
1. See summaries only (recommended)
2. Filter to specific types (YouTube, articles, etc.)
3. Search for a more specific AI topic"
`;
