gemini-tavily-search

Purpose

Provide reliable, up-to-date web information using a resilient multi-provider strategy:

Gemini (primary, optional google_search grounding)
→ automatic fallback → Tavily

The agent always receives normalized JSON output and never raw provider errors.

Activation Criteria

Use this skill ONLY when:

• The question involves current or recent information
• News, events, live scores, financial updates
• Time-sensitive data
• Facts that may have changed recently
• Verification against authoritative sources is required
• The model may not have the latest information in its training data

Do NOT Activate When

Do NOT use this skill when:

• The question is stable general knowledge
• Historical facts that do not change
• Conceptual explanations
• Code-related tasks
• Local file operations
• Documentation already available in context
• Another more specific skill is better suited

Internal Logic

1. Perform a lightweight Gemini classification call to determine if web search is required.
2. If web search is NOT required → answer directly via Gemini without tools.
3. If web search IS required → call Gemini with google_search tool enabled.
4. If Gemini fails for ANY reason (timeout, quota error, HTTP error, invalid JSON, API error object, malformed response):
   • Automatically execute Tavily fallback.
5. Normalize provider output into unified JSON schema.
6. Always return valid structured JSON.

The agent must not describe fallback logic to the user.

Input

Call the script with a single JSON argument.

Required

• query (string)

Optional (forwarded to Tavily if fallback occurs)

• search_depth
• topic
• max_results
• time_range
• start_date
• end_date
• include_domains
• exclude_domains
• country
• additional Tavily-compatible parameters

Environment Requirements

Required:

• TAVILY_API_KEY
• GEMINI_API_KEY

Optional:

• GEMINI_MODEL (default: gemini-2.5-flash-lite)

{
  "env": {
    "GEMINI_MODEL": "gemini-2.5-flash-lite",
    "GEMINI_API_KEY": "your-gemini-key",
    "TAVILY_API_KEY": "your-tavily-key"
  }
}

Output Schema (Unified)

The tool always returns JSON in this format:

{
  "provider": "gemini | tavily",
  "answer": "text or null",
  "results": [
    {
      "title": "Source title",
      "url": "https://example.com",
      "snippet": "Relevant excerpt"
    }
  ],
  "fallback": true | false
}

If both providers fail:

{
  "provider": "tavily",
  "answer": null,
  "results": [],
  "fallback": true,
  "error": "tavily_failed"
}

Post-Processing Requirement

After generating the final answer for the user, append:

Searched with: \x3Cprovider>

Where \x3Cprovider> matches the returned JSON:

• "gemini"
• "tavily"

If no web search was used:

Searched with: none