Alerts agent based on GEE LCL layers

[yellowcap]

An agent to feed the first alerts workflow. The agent needs to

• [x] Find GADM areas based on bbox or text
• [x] Compute DIST alert statistics for the area
• [x] Create a text description of the result

Based on LCL layers on GEE.

[yellowcap]

Som insight on getting geojson from tools: this is a lot of tokens for llms, but its hard to shield the llm from tool output.

It's not possible to update state from within a tool. We need to find a way to pass larger amounts of data from tools to the frontend without giving it to the LLM for interpretation. We probably want to pass part of the data as-is to the frontend, and part to the LLM for interpretation.

We can inject the state into the tool using InjectedState , but not update the state (or at least its highly discouraged).

For instance, pass disturbance statistics data to LLM, but send vectorized disturbances directly to frontend. I have not figured out how to solve that, and the vectorized geojson generates quite a lot of tokens.

Related discussions

https://github.com/langchain-ai/langgraph/discussions/1616 https://github.com/langchain-ai/langgraph/discussions/337

[yellowcap]

Ok this was a bit hard to figure out, but there is a simple solution to this actually! 🫏

The @tool decorator has an additional argument: response_format (Literal['content', 'content_and_artifact']).

If this is is et to content_and_artifact, any kind of data can be passed to downstream as artifacts, and they will not be passed to the model. We can leverage this to feed data to the user.