	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > Gemini Dynamic Retrieval and Grounding with Google Search

Gemini Dynamic Retrieval and Grounding with Google Search

Author: Venkata Sudhakar

Gemini grounding with Google Search connects the model to live web data at inference time. When grounding is enabled, Gemini automatically decides whether to search the web before answering. For questions about current events, live prices, recent releases, or fast-changing facts, it searches Google and grounds the response in real sources with citations. For stable factual questions it answers from its own training knowledge without searching. This dynamic decision saves unnecessary search calls while ensuring the model never fabricates an answer it cannot verify.

Dynamic retrieval is the configuration that controls when Gemini searches versus when it answers from training knowledge. The key parameter is dynamic_threshold: a value from 0.0 to 1.0 where lower values mean more aggressive searching and higher values mean more conservative searching. Setting threshold to 0.0 forces search on every query. Setting it to 1.0 disables search entirely. The recommended starting point is 0.3 for consumer applications and 0.5 for enterprise internal knowledge applications where you want fewer external searches.

The below example builds a market intelligence agent for a retail company that uses dynamic retrieval to answer both current market questions (grounded in live search) and internal policy questions (answered from training), with full source citations for all searched answers.

from google import genai
from google.genai import types

client = genai.Client(api_key="your-gemini-api-key")

def ask_with_grounding(question: str, threshold: float = 0.3) -> dict:
    resp = client.models.generate_content(
        model="gemini-2.0-flash",
        config=types.GenerateContentConfig(
            system_instruction=(
                "You are a market intelligence analyst for ShopMax India. "
                "Answer questions about the electronics market and retail industry. "
                "Always cite your sources when using web search results. "
                "Be specific with data points and dates."
            ),
            tools=[types.Tool(
                google_search=types.GoogleSearch(
                    dynamic_retrieval_config=types.DynamicRetrievalConfig(
                        mode=types.DynamicRetrievalConfigMode.MODE_DYNAMIC,
                        dynamic_threshold=threshold
                    )
                )
            )],
            max_output_tokens=400
        ),
        contents=[question]
    )
    candidate = resp.candidates[0]
    answer    = candidate.content.parts[0].text
    grounding = candidate.grounding_metadata
    sources   = []
    if grounding and grounding.grounding_chunks:
        for chunk in grounding.grounding_chunks:
            if hasattr(chunk, "web") and chunk.web:
                sources.append({"title": chunk.web.title,
                                "uri":   chunk.web.uri})
    searched = bool(sources)
    return {"answer": answer, "searched": searched, "sources": sources}

Testing dynamic retrieval on current market questions versus stable factual questions,

It gives the following output showing dynamic retrieval in action,

Q: What is the current market share of Samsung in the Indian smartphone market?
Expected: CURRENT - expect search
Searched: True
Answer: According to IDC Q1 2025 data, Samsung holds approximately 18% market
        share in India, placing it second after Vivo at 19%. Samsung strengthened
        its position with the Galaxy S25 series launch in February 2025.
Sources:
  - IDC India Smartphone Market Q1 2025 - IDC.com
  - Samsung Galaxy Market Share India 2025 - CounterPoint Research

Q: What are the four Ps of marketing?
Expected: STABLE - expect no search
Searched: False
Answer: The four Ps of marketing are Product (what you sell), Price (what you
        charge), Place (where you sell it), and Promotion (how you communicate
        value). This framework was introduced by E. Jerome McCarthy in 1960.

Q: What is the GST rate on mobile phones in India currently?
Expected: POLICY - may search to verify current rate
Searched: True
Answer: Mobile phones in India attract GST at 12% as of 2025. This rate
        applies to smartphones and feature phones under HSN code 8517.
Sources:
  - GST Rate on Mobile Phones 2025 - ClearTax

# Dynamic threshold 0.3: searched for current facts, did not search for stable theory
# Source citations included automatically for all searched answers
# Zero hallucination on current market data - all grounded in live web results

Threshold comparison showing how the value controls search aggressiveness,

Threshold 0.0 | Searched: True  | Sources: 4  (always searches)
Threshold 0.3 | Searched: True  | Sources: 3  (searches for current topics)
Threshold 0.7 | Searched: False | Sources: 0  (conservative - answers from knowledge)
Threshold 1.0 | Searched: False | Sources: 0  (never searches)

# 0.0: maximum freshness - use for news and market intelligence agents
# 0.3: balanced - use for general enterprise assistants
# 0.7: conservative - use when internal knowledge base is primary source
# 1.0: no search - equivalent to disabling grounding entirely

Grounding production best practices: start with threshold 0.3 and monitor your search rate in Cloud Logging. If searches happen too often for clearly stable questions, raise to 0.5. If the agent is missing current data it should have searched for, lower to 0.2. Always display citations to users when sources are present - it builds trust and allows users to verify facts. For internal enterprise agents where you want to search only your internal knowledge base rather than the public web, use Vertex AI Search grounding (Tutorial 330) instead of Google Search grounding. Combine dynamic retrieval with your custom tools for a hybrid agent: custom tools for your proprietary data, Google Search grounding for current external market data.

Send your comments, suggestions or queries regarding this site to [email protected].