Parsing JSON Responses from Gemini AI

Gemini often wraps JSON responses in markdown code blocks that need to be stripped.

Problem

Gemini responses may look like:

Here's the translation:

```json
[
  {"original": "日本語", "translated": "Japanese"}
]
```

Solution: Strip Markdown Code Blocks

function parseGeminiJSON(text: string): any {
  // Remove markdown code blocks with optional language identifier
  const cleaned = text.replace(/```(?:json)?\s*([\s\S]*?)\s*```/, '$1');

  // Parse the cleaned JSON
  return JSON.parse(cleaned.trim());
}

Regex Breakdown

• ``` - Matches opening code block (triple backticks)
• (?:json)? - Optional "json" language identifier
• \s* - Optional whitespace
• ([\s\S]*?) - Captures content (non-greedy)
• \s* - Optional whitespace
• ``` - Matches closing code block

Robust Version

function parseGeminiJSON(text: string): any {
  let cleaned = text;

  // Remove markdown code blocks
  cleaned = cleaned.replace(/```(?:json)?\s*([\s\S]*?)\s*```/, '$1');

  // Remove any text before/after JSON
  const jsonMatch = cleaned.match(/\[[\s\S]*\]|\{[\s\S]*\}/);
  if (jsonMatch) {
    cleaned = jsonMatch[0];
  }

  return JSON.parse(cleaned.trim());
}

Usage Example

const response = await model.generateContent(prompt);
const rawText = response.response.text();

try {
  const parsed = parseGeminiJSON(rawText);
  console.log(parsed);
} catch (error) {
  console.error('Failed to parse JSON:', rawText);
}

Common Patterns