agentic_rag

Google ADK Agentic RAG with Comet Opik

Trace a Google ADK router agent with Comet Opik.

What this does

This example runs a Google ADK router agent that decides whether each query should use local RAG over the IMF Financial Access Survey PDF or live web search. It indexes the PDF into a local Qdrant database, exposes retrieve_docs and web_search tools, and traces the routing decision and tool calls with Opik.

Prerequisites

This is a uv project — dependencies live in pyproject.toml.

uv sync

Copy .env.example to .env (or export the variables). With GOOGLE_API_KEY / Opik credentials unset, the example runs in DRY_RUN and prints what it would do instead of calling Gemini/Opik.

Variable	Required	Description
GOOGLE_API_KEY	for a live run	Gemini API key (Google AI Studio). Unset → DRY_RUN.
OPIK_API_KEY	for a live run	Opik API key from comet.com/opik. Unset → DRY_RUN.
OPIK_WORKSPACE	for a live run	Your Opik workspace. Unset → DRY_RUN.
OPIK_PROJECT_NAME	no	Project traces are logged to (default google-adk-rag).
GADK_MODEL	no	Gemini model the router runs on (default gemini-2.5-flash).

Running it

uv run python index.py   # build the local Qdrant index from the IMF PDF
uv run python main.py    # run the traced ADK router

# or both, the way CI does:
bash run.sh

How it works

1. Index the PDF — index.py downloads the IMF report, chunks it, embeds it, and writes vectors into the local db directory.
2. Expose tools — tools.py defines retrieve_docs for Qdrant search and web_search for DuckDuckGo search.
3. Run the router — main.py creates the router_agent, traces it with OpikTracer (logging to OPIK_PROJECT_NAME, default google-adk-rag), and runs one query through ADK.