​
Why Use Code Reranker?

Code Reranker revolutionizes how we handle large codebases with LLMs:

  • No Embeddings Required: Direct semantic understanding of code relationships
  • Context Optimization: Intelligently selects the most relevant code within token limits
  • Code-Specific Accuracy: Outperforms generalist rerankers by understanding code structure, dependencies, and programming patterns

Unlike general-purpose rerankers, Code Reranker is specifically trained on code relationships and programming patterns, making it significantly more accurate for code-related tasks.

​
Performance Metrics

Code Reranker v2 additionally provides relevance scores that can be used to manually optimize token usage. The plot below shows how different threshold values affect performance:

​
Understanding the Metrics

  • Threshold: The minimum relevance score required for a file to be included
  • Recall: Percentage of relevant files that are included (higher is better)
  • Tokens Saved: Percentage of tokens that are excluded
  • High Recall (95%+): Use threshold ≤ 0.08
  • Balanced (80-90%): Use threshold between 0.20-0.30

​
Example Request

import requests

url = "https://ranker.endpoint.relace.run/v2/code/rank"  # Using v2 endpoint
headers = { "Content-Type": "application/json", "Authorization": "Bearer [YOUR_API_KEY]" }

query = "Optimize the search function for better performance with large arrays"
codebase = [ { "filename": "src/search.ts", "code": "function findItem(array: Item[], targetId: string): Item | undefined {\\n  for (let i = 0; i < array.length; i++) {\\n    const item = array[i];\\n    if (item.id === targetId) {\\n      return item;\\n}\\n}\\n  return undefined;\\n}"}, { "filename": "src/types.ts", "code": "interface Item {\\n  id: string;\\n  value: string;\\n  metadata?: Record<string, unknown>;\\n}"}, { "filename": "src/render.ts", "code": "function renderItems(items: Item[]): void {\\n  const container = document.getElementById('items-container');\\n  if (!container) return;\\n  \\n  container.innerHTML = '';\\n  \\n  for (const item of items) {\\n    const element = document.createElement('div');\\n    element.textContent = item.value;\\n    element.dataset.id = item.id;\\n    container.appendChild(element);\\n}\\n}"}, { "filename": "src/styles.css", "code": ".container {\\n  max-width: 1200px;\\n  margin: 0 auto;\\n  padding: 20px;\\n}\\n\\n.header {\\n  background: #f5f5f5;\\n  padding: 1rem;\\n}"}, { "filename": "src/constants.ts", "code": "export const APP_NAME = 'MyApp';\\nexport const VERSION = '1.0.0';\\nexport const MAX_ITEMS = 1000;"}
]

data = { "query": query, "codebase": codebase, "token_limit": 128000 }

response = requests.post(url, headers=headers, json=data)
ranked_code = response.json()

​
Example Response

{
  "results": [
    {
      "filename": "src/search.ts",
      "score": 0.953125
    },
    {
      "filename": "src/utils.ts",
      "score": 0.234375
    }
  ],
  "numTokens": 261
}