Summary
An agent can run a proper sweep instead of a single search.
Deep search issues several rounds of queries, reads across them and returns one grounded answer; quick search stays there for the questions that only need a look. The depth is a parameter, so the cost of a run is something you decide.
What it is
A research agent behind three calls, built on LangGraph and Gemini. Deep search runs multiple search iterations and folds the results together; quick search is a single pass. Citations come from Gemini's grounding metadata rather than being assembled afterwards.
What you get
- A multi-round investigation of a topic, with the iteration count and results per query under your control
- A single-pass search for questions that do not need rounds
- Search language selectable per call
- Source citations attached automatically, from grounding metadata
- The state of the configured search tools, readable before you depend on them
Requirements
Python 3.10 or newer, uv, and a Gemini API key in the environment. Runs via uvx from the cloned project. The documentation is in Japanese.
Setup effort
One command plus a key — uvx --from . langgraph-deep-search, then supply credentials