The latest stable version of VS Code is 1.96.4 (as of January 2026)...

[citation:1] ← Citation source marker

**You should see**:
- Model response includes citation sources (`[citation:1]`, etc.)
- Response content is up-to-date, not old version from training data

### Step 5: Test Different Thresholds

Try adjusting `grounding_threshold` and observe model behavior changes:

```json
// Low threshold (frequent searches)
"grounding_threshold": 0.1

// High threshold (rare searches)
"grounding_threshold": 0.7

After each adjustment, test with the same question and observe:

• Whether it searches (check if response has citations)
• Search count (multiple citations)
• Response speed

You should see:

• Low threshold: More frequent searches, but slightly slower response
• High threshold: Fewer searches, but potentially inaccurate answers

Checkpoint ✅

Click to expand verification checklist

Complete the following checks to confirm correct configuration:

• [ ] Configuration file contains web_search configuration
• [ ] default_mode set to "auto"
• [ ] grounding_threshold is between 0.0 and 1.0
• [ ] Make a request requiring real-time information, model returns response with citations
• [ ] After adjusting threshold, model search behavior changes

If all pass, Google Search Grounding is correctly enabled!

Troubleshooting

Problem 1: Model Not Searching

Symptoms: After enabling auto mode, the model still doesn't search, and no citation sources appear.

Causes:

• Threshold too high (e.g., 0.9), model needs very high confidence to search
• Question itself doesn't require searching (e.g., programming problems)

Solutions:

• Lower grounding_threshold to 0.2 or below
• Test with questions that clearly need real-time information (e.g., "What's the weather today", "Latest news")

Problem 2: Too Frequent Searches, Slow Responses

Symptoms: Every question triggers a search, response time significantly increases.

Causes:

• Threshold too low (e.g., 0.1), model triggers searches too frequently
• Question type itself requires real-time information (e.g., stock prices, news)

Solutions:

• Raise grounding_threshold to 0.5 or higher
• If the task doesn't need real-time information, change default_mode to "off"

Problem 3: Configuration File Format Error

Symptoms: Plugin reports error, cannot load configuration.

Cause: JSON format error (e.g., extra commas, mismatched quotes).

Solution: Use a JSON validation tool to check the configuration file, ensure correct format.

# Validate JSON format
cat ~/.config/opencode/antigravity.json | python3 -m json.tool

Summary

• Google Search Grounding enables Gemini models to search real-time web information
• Enable with web_search.default_mode: "auto", disable with "off"
• grounding_threshold controls search frequency: lower values mean more frequent searches
• Default threshold 0.3 works for most scenarios, adjust based on experience
• Model will cite sources in responses, marked as [citation:1], etc.

Next Lesson Preview

Next, we'll learn about Dual Quota System.

You'll learn:

• How the two independent quota pools (Antigravity and Gemini CLI) work
• How to switch between quota pools to maximize utilization
• Best practices for quota pooling

---

Appendix: Source Code Reference

Click to view source code locations

Last updated: 2026-01-23

Feature	File Path	Lines
Google Search Config Schema	src/plugin/config/schema.ts	303-319
Google Search Type Definitions	src/plugin/transform/types.ts	85-88
Google Search Injection Logic	src/plugin/transform/gemini.ts	402-419
Google Search Config Loading	src/plugin/config/loader.ts	173-184
Google Search Config Application	src/plugin.ts	1194-1196

Key Configuration Options:

• web_search.default_mode: "auto" or "off", defaults to "off"
• web_search.grounding_threshold: 0.0 - 1.0, defaults to 0.3

Key Functions:

• applyGeminiTransforms(): Applies all Gemini transformations, including Google Search injection
• normalizeGeminiTools(): Normalizes tool format, preserves googleSearchRetrieval tool
• wrapToolsAsFunctionDeclarations(): Wraps tools as functionDeclarations format

How It Works:

1. Plugin reads web_search.default_mode and web_search.grounding_threshold from configuration file
2. When mode === "auto", injects googleSearchRetrieval tool in the request's tools array:

   {
     "googleSearchRetrieval": {
       "dynamicRetrievalConfig": {
         "mode": "MODE_DYNAMIC",
         "dynamicThreshold": 0.3  // grounding_threshold
       }
     }
   }

3. Gemini model decides whether to call the search tool based on the threshold
4. Search results are included in the response, marked with citation sources ([citation:1], etc.)