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.
The gemini-file-search-demo repository is organised around a progression: you start with a minimal SDK agent that has no RAG, then layer in File Search, and finally refactor everything into the Agent Development Kit (ADK) using the Agent-as-a-Tool pattern. Understanding the layout up front makes it easier to follow each step of the codelab.
Directory tree
gemini-file-search-demo/
├── app/
│ ├── basic_agent_adk/ # Agent with Google Search, using ADK framework
│ │ └── agent.py
│ ├── rag_agent_adk/ # Agent with Google Search and File Search, using ADK framework
│ │ ├── agent.py
│ │ └── tools_custom.py
│ ├── sdk_agent.py # Agent using GenAI SDK (no ADK) with Google Search tool
│ └── sdk_rag_agent.py # Agent using GenAI SDK (no ADK) with Gemini File Search tool
├── data/
│ └── story.md # Sample story with "bespoke content" to use with Gemini File Search Store
├── notebooks/
│ └── file_search_store.ipynb # Jupyter notebook for creating and managing Gemini File Search Store
├── .env.template # Template for environment variables — make a copy as .env
├── Makefile # Makefile for make commands
├── pyproject.toml # Project configuration and dependencies
└── README.md
Agent files
Each agent in app/ represents a distinct implementation approach. They are designed to be compared side by side.
app/sdk_agent.py — SDK baseline (Google Search only)
The simplest possible Gemini agent. It instantiates a genai.Client, attaches types.GoogleSearch() as a tool, and runs a REPL loop. There is no RAG. Its purpose is to establish a baseline and demonstrate the model’s inability to answer questions about bespoke content that is not publicly indexed.
app/sdk_rag_agent.py — SDK agent with File Search
A direct extension of the baseline that replaces Google Search with the Gemini File Search tool. It looks up a File Search Store by the display name stored in STORE_NAME, attaches it via types.FileSearch(file_search_store_names=[store.name]), and falls back gracefully when the store is not found, logging a warning and disabling RAG rather than crashing.
app/basic_agent_adk/agent.py — ADK agent (Google Search only)
Implements the same Google Search capability as the baseline, but using the ADK Agent class. It introduces the Agent-as-a-Tool pattern: a SearchAgent sub-agent owns the google_search tool, and a root_agent delegates to it via AgentTool(agent=search_agent). A strict “fail fast” system instruction prevents infinite search retries on topics the model cannot find.
app/rag_agent_adk/agent.py — ADK agent with File Search and Google Search
The complete multi-agent solution. A root_agent orchestrates two specialist sub-agents — a SearchAgent for general knowledge and a RagAgent for bespoke knowledge — each wrapped as an AgentTool. The root agent’s instruction directs it to try RagAgent first. This separation is required because the Gemini API does not allow Google Search and File Search in the same single-agent request.
Because ADK does not yet include a built-in wrapper for the File Search tool, this module provides one. FileSearchTool extends BaseTool and implements process_llm_request to inject types.FileSearch configuration directly into the outgoing model request before it is sent. It is used exclusively by RagAgent.
Data and notebooks
| Path | Purpose |
|---|
data/story.md | The sample story “The Wormhole Incursion” — bespoke content that Gemini has no training knowledge of. Used to demonstrate and verify RAG. |
notebooks/file_search_store.ipynb | Interactive notebook for creating a File Search Store, uploading files, and inspecting the store’s contents. Run this before the RAG agents. |
Configuration files
| File | Purpose |
|---|
.env.template | Template for the three required environment variables. Copy to .env and fill in your values. |
Makefile | Shorthand make targets for common tasks: installing dependencies, running each agent, launching the ADK web UI, and running linters. |
pyproject.toml | Project metadata, Python version constraint, and all dependencies managed by uv. |
Key dependencies
The project requires Python >=3.12. Core dependencies declared in pyproject.toml:
| Package | Role |
|---|
google-genai | Google Gen AI SDK — the low-level client used by the SDK agents |
google-adk | Agent Development Kit — Agent, AgentTool, BaseTool, built-in tools |
google-cloud-aiplatform[adk] | Vertex AI platform support with ADK extras |
fastapi | ASGI web framework (used internally by ADK’s web server) |
uvicorn | ASGI server that backs the ADK web UI |
jupyter — jupyter and ipython, needed to run notebooks/file_search_store.ipynblint — ruff, mypy, codespell, plus type stubs, used by make lint
