Ground the agent in your own documents

An agent is only as good as what it was given to read. In most frameworks the retrieval step disappears into SDK code: you cannot see which chunks were fetched for a given answer, so when the answer is wrong you are debugging a black box. FluidGrids takes the opposite approach and makes retrieval a node like every other node.

What is on the canvas today is a full vector-store package: create, list, delete, and inspect collections; ingest raw chunks or whole documents; generate embeddings; run similarity search; run hybrid search. It binds two connections — one for the vector store, one for the embedding provider — and both are resolved server-side by the executor at run time, exactly like any other connector credential.

So the honest pattern today is retrieval as an explicit step. A trigger fires, a search node pulls the relevant passages, a set node assembles the grounded prompt, the AI Agent node answers with its chat model attached, a structured parser pins the output to a schema, and a destination node delivers it. Every one of those is a node state in run history. When an answer looks wrong, you open the run and read what was retrieved and what the model was handed — which is a stronger audit position than a framework that hides the same step. Because the vector-store operations also expose themselves as agent tools, an agent can alternatively call search itself as a tool, with the executor performing the call.

Here is what is not wired, precisely. The node declares a vector-store adapter output type and ships a supply-vector-store operation, but the executor's adapter map currently routes only chat model, memory, tool, and output parser. Attach a vector store to the agent and the agent does not consume it as a retrieval port. Verifying and deepening that composition — vector store as a first-class agent port, memory adapters proven end to end, and an adapter for calling external MCP tool servers — is tracked on our roadmap and is not built.

And one thing we will not claim at all: quality. There is no evaluation set, no accuracy figure, no hallucination rate. An evaluation dataset with a CI gate on model and prompt changes is on our roadmap, and answer citations with a hallucination disclaimer sit on our roadmap. Until those land, treat retrieval here as a step you can inspect and version — which is real and useful — rather than as a guarantee about what the model concludes from it.