Documentation Index
Fetch the complete documentation index at: https://mintlify.com/derailed-dash/gemini-file-search-demo/llms.txt
Use this file to discover all available pages before exploring further.
With a populated File Search Store in place, the next step is to connect it to an agent. app/sdk_rag_agent.py is structurally similar to the basic SDK agent, but instead of attaching a Google Search tool it retrieves your File Search Store by name and passes it as a FileSearch tool. The result is an agent that can answer questions about your bespoke content accurately — using only the documents you’ve uploaded — without any manual chunking, embedding, or vector database work.
What changed from the basic agent
The basic agent used:
google_search_tool = types.Tool(google_search=types.GoogleSearch())
file_search_tool = types.Tool(file_search=types.FileSearch(file_search_store_names=[store.name]))
genai.Client() initialization, the client.chats.create(...) call, and the REPL loop — follows the same pattern.
The full agent code
"""
Native SDK RAG Agent with Gemini File Search Tool
"""
import logging
import os
from dotenv import load_dotenv
from google import genai
from google.genai import types
# Configure logging
logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
logger = logging.getLogger(__name__)
load_dotenv()
STORE_NAME = os.getenv("STORE_NAME")
logger.info(f"Store name: {STORE_NAME}")
model = os.getenv("MODEL", "gemini-2.5-flash")
logger.info(f"Using model: {model}")
def get_store(client: genai.Client, store_name: str) -> types.FileSearchStore | None:
"""Retrieve a store by display name"""
try:
for a_store in client.file_search_stores.list():
if a_store.display_name == store_name:
return a_store
except Exception as e:
logger.error(f"Error listing stores: {e}")
return None
def main() -> None:
"""
Runs a simple interactive chat with Gemini using the File Search Tool.
"""
# 1. Initialize the client
client = genai.Client()
# 2. Retrieve the File Search Store
logger.info(f"Looking for File Search Store: {STORE_NAME}...")
store = None
if STORE_NAME:
store = get_store(client, STORE_NAME)
else:
logger.warning("STORE_NAME env var not set. Skipping store lookup.")
# 3. Build Tools List
tools_list = []
if store and store.name:
logger.info(f"Found store: {store.name}")
file_search_tool = types.Tool(file_search=types.FileSearch(file_search_store_names=[store.name]))
tools_list.append(file_search_tool)
else:
logger.warning(f"Store '{STORE_NAME}' not found! RAG capabilities disabled.")
# 4. Create the chat session
chat = client.chats.create(
model=model,
config=types.GenerateContentConfig(
tools=tools_list,
automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=False),
temperature=0.7,
system_instruction="""You are a helpful assistant with access to a story knowledge base (via File Search).""",
),
)
if store:
print(f"Connected to Store: {STORE_NAME}")
print("Type 'exit' or 'quit' to stop.")
print("------------------------------------------")
while True:
try:
user_input = input("User: ").strip()
if user_input.lower() in ("exit", "quit"):
print("Goodbye!")
break
if not user_input:
continue
# 4. Send message and get response
response = chat.send_message(user_input)
# 5. Print response
if response.text:
print(f"Agent: {response.text}")
# Print citations if available to prove RAG is working
if response.candidates and response.candidates[0].grounding_metadata:
gm = response.candidates[0].grounding_metadata
if gm.grounding_chunks:
print(f"\n[Grounding] Found {len(gm.grounding_chunks)} chunks from File Search.")
if gm.search_entry_point:
print(f"[Grounding] Search Entry Point: {gm.search_entry_point.rendered_content}")
else:
print("Agent: [No text response]")
except Exception as e:
logger.error(f"An error occurred: {e}")
print(f"Error: {e}")
if __name__ == "__main__":
main()
Code walkthrough
Retrieving the store with get_store
def get_store(client: genai.Client, store_name: str) -> types.FileSearchStore | None:
"""Retrieve a store by display name"""
try:
for a_store in client.file_search_stores.list():
if a_store.display_name == store_name:
return a_store
except Exception as e:
logger.error(f"Error listing stores: {e}")
return None
client.file_search_stores.list() returns all stores in your project. The function matches on display_name — the human-readable name you set in your .env file as STORE_NAME. It returns the FileSearchStore object (which carries the resource name needed to build the tool) or None if no match is found.
file_search_tool = types.Tool(file_search=types.FileSearch(file_search_store_names=[store.name]))
tools_list.append(file_search_tool)
store.name is the unique resource identifier (e.g., filestores/abc123), not the display name. Passing it inside file_search_store_names tells Gemini which store to search when it needs to ground a response.
If get_store returns None (because the store was not found), tools_list stays empty and the agent starts without RAG. A warning is logged so the problem is visible immediately.
After each response, the agent inspects response.candidates[0].grounding_metadata and prints two pieces of diagnostic information:
if gm.grounding_chunks:
print(f"\n[Grounding] Found {len(gm.grounding_chunks)} chunks from File Search.")
if gm.search_entry_point:
print(f"[Grounding] Search Entry Point: {gm.search_entry_point.rendered_content}")
- Grounding chunks: the number of document chunks the File Search tool retrieved and used to construct the answer. A non-zero count confirms RAG is active.
- Search entry point: the rendered query Gemini issued against your store, useful for understanding how the model interpreted the question.
Running the agent
Demo walkthrough
Confirm the store is connected
When the agent starts, look for the confirmation line:Connected to Store: demo-file-store
STORE_NAME in your .env matches the display name you used when creating the store in the notebook. Ask the question that previously failed
Ask the same question the basic agent could not answer:> Who pilots the 'Too Many Pies' ship?
data/story.md, and returns the correct, story-specific answer. Inspect the grounding metadata
After the answer you’ll see output like:[Grounding] Found 5 chunks from File Search.
[Grounding] Search Entry Point: ...
Key insight: STORE_NAME must match the display name
The STORE_NAME environment variable must exactly match the display_name you gave the store when you created it in the notebook. If the names differ, get_store returns None, RAG is silently disabled, and the agent behaves like the basic search-only version.
get_store function looks up stores by display_name, not by the internal resource name. As long as your .env value and the notebook creation cell use the same string, the lookup will succeed every time.
What’s next
This agent uses File Search in place of Google Search. If you need both — answering questions from your documents and from the web — the two tools cannot be attached to the same agent in a single request. The solution is the Agent-as-a-Tool pattern used in the ADK agents covered in the next section of this codelab.