Grounded Search

Torale uses Gemini's grounded search to combine Google Search with AI evaluation.

What is Grounded Search?

Grounded search means the AI's responses are grounded in actual search results, not just the model's training data. This provides:

• Current information - Real-time web data
• Source attribution - Verifiable citations
• Reduced hallucinations - Answers based on search results
• Temporal awareness - Knows about recent events

How It Works

1. Search Phase

# Gemini queries Google Search with user's query
search_results = await gemini.search_with_grounding(
    query="When is the next iPhone being released?",
    search_provider="google"
)

2. Answer Extraction

# LLM synthesizes answer from search results
answer = await gemini.generate(
    prompt=f"Based on these search results, {query}",
    results=search_results,
    max_length="2-4 sentences"  # Concise for emails
)

3. Condition Evaluation

# LLM determines if condition is met
evaluation = await gemini.evaluate(
    answer=answer,
    condition="A specific release date has been announced",
    reasoning=True
)

4. Source Extraction

# Extract and filter grounding citations
sources = extract_sources(search_results)
sources = filter_infrastructure_urls(sources)  # Remove Vertex AI URLs

Implementation

Executor Code

class GroundedSearchExecutor:
    async def execute(self, config: dict) -> dict:
        query = config["search_query"]
        condition = config["condition_description"]

        # Perform grounded search
        result = await self.gemini_client.search(
            query=query,
            grounding_config={
                "provider": "google",
                "dynamic_retrieval": True
            }
        )

        # Extract answer
        answer = result.text

        # Evaluate condition
        evaluation = await self.evaluate_condition(
            answer=answer,
            condition=condition
        )

        # Extract sources
        sources = self.extract_grounding_sources(result)

        return {
            "condition_met": evaluation.is_met,
            "answer": answer,
            "reasoning": evaluation.reasoning,
            "sources": sources
        }

Source Filtering

Gemini sometimes includes Vertex AI infrastructure URLs in grounding sources. We filter these:

def filter_infrastructure_urls(sources):
    """Remove Vertex AI infrastructure URLs"""
    filtered = []
    for source in sources:
        # Skip Vertex AI redirect URLs
        if "vertexaisearch.cloud.google.com" in source["uri"]:
            continue
        if "cloudfront.net" in source["uri"]:
            continue
        filtered.append(source)
    return filtered

Result: Clean source URLs like apple.com instead of redirect URLs.

Temporal Context

The executor provides the LLM with temporal context to improve change detection:

async def execute_with_context(self, task):
    last_execution_time = task.last_executed_at
    time_since_last = datetime.now() - last_execution_time

    prompt = f"""
    Query: {task.search_query}
    Condition: {task.condition_description}

    Context: This task was last checked {time_since_last} ago.
    Focus on NEW information since then.
    """

Benefits:

• Reduces false positives from old news
• Improves change detection accuracy
• Helps LLM identify meaningful updates

Response Format

Successful Evaluation

{
  "condition_met": true,
  "answer": "Apple has announced iPhone 16 will be released on September 20, 2024. Pre-orders begin September 13.",
  "reasoning": "The condition is met because Apple's official announcement confirms a specific release date of September 20, 2024.",
  "sources": [
    {
      "uri": "https://www.apple.com/newsroom/...",
      "title": "Apple announces iPhone 16"
    },
    {
      "uri": "https://www.theverge.com/...",
      "title": "iPhone 16 release date confirmed"
    }
  ]
}

Condition Not Met

{
  "condition_met": false,
  "answer": "No official iPhone 16 release date has been announced yet. Rumors suggest a fall 2024 release.",
  "reasoning": "The condition requires an official announcement, but current information consists only of rumors and speculation.",
  "sources": [
    {
      "uri": "https://www.macrumors.com/...",
      "title": "iPhone 16 rumored for fall release"
    }
  ]
}

Model Configuration

Primary Model

{
  "model": "gemini-2.0-flash-exp",
  "search_provider": "google",
  "temperature": 0.7,
  "max_output_tokens": 512
}

Why Gemini Flash:

• Native Google Search integration
• Fast response times (2-5 seconds)
• Cost-effective (~$0.001 per execution)
• Optimized for search tasks

Fallback Models

If Gemini unavailable:

1. OpenAI GPT-4 Turbo + web search
2. Anthropic Claude 3 + web scraping

Performance

Typical execution:

• Search: 1-2 seconds
• Answer generation: 1-2 seconds
• Condition evaluation: 0.5-1 second
• Total: 3-5 seconds

Best Practices

Query Writing

• Be specific about what you're searching for
• Include relevant context (product names, locations)
• Write as natural questions

Condition Writing

• Make objectively evaluable
• Include specific criteria (dates, prices, availability)
• Avoid subjective terms

Next Steps

• Learn about State Tracking
• Understand Temporal Workflows
• View Executor System