pyflowred-node-ai-google

Google Gemini for PyFlowRED's AI nodes.

pip install pyflowred-node-ai-google

or, in a running instance, Menu → Manage palette → Install → pyflowred-node-ai-google, which also declares it so a rebuilt instance installs it again by itself.

What it adds

One configuration node, ai-provider-google. PyFlowRED's ai chat, ai structured, ai agent, ai embed and ai query point at it and ask it for a model; everything Google-specific lives here.

The same model, two backends

langchain-google-genai carries both Google backends in one class, so the Authentication dropdown is the whole difference between them:

Moving a flow from a key on a laptop to a service account in a cluster is that dropdown and nothing else: not a different node, not a different package, not a different model name.

The credential is not in the flow file

It is a credential property, so it lands encrypted in the instance's credentials file. Written as ${secret:NAME} it is fetched from the instance's secrets provider when the call is made, so a rotated key is picked up without a deploy and neither file ever holds the value.

Thinking is the biggest number on the bill

Measured against gemini-3.8-flash: the one-word answer "Paris" cost 52 output tokens, of which 51 were thinking — and 3 with thinking set to off. For a flow answering a thousand messages a day that is the difference between the bill and ten times the bill. It is a field on this node, and a function node or a message may override it for one call.

What a message may override

model, temperature, maxTokens, topP, thinking, dimensions, taskType and extra — through the function node's own fields or through msg.ai. A credential never: it belongs to the connection, not to the operation.

Independent on purpose

This package depends on langchain-google-genai and not on PyFlowRED, and it imports nothing from it. The contract a provider satisfies is three attributes on an object, so the engine and the providers are released apart. See docs/writing-nodes.md in PyFlowRED for how to write another one.

Apache-2.0.