Griptape Tool: RAG

The RAG tool is the "I'm building my own retrieval pipeline" node - the one that sits in the middle between "just use a prebuilt knowledge base" and "I want to control every stage." It wraps Griptape's RagTool, and what makes it different from the simpler vector-store tools in this pack is the input it demands: a RAG_ENGINE. That engine is a full retrieval + response pipeline you assemble from the pack's other RAG module nodes (query transformation, retrieval from a vector store, reranking, response generation). If you've graduated past drop-in RAG and want to tune how retrieval actually works, this is where the pack gets serious.

The name "RAG" is doing a lot here - this is the tool that turns a RAG engine into something an agent can actually call. You describe what the information is about (the description is what the agent reads to decide whether to use this tool), you hand it a built engine, and the agent gets grounded answers from whatever retrieval you configured. The other useful detail: this node returns a second output, a RULES ruleset, which bakes in instructions like "include all provided footnotes" - because Griptape RAG engines can attach footnotes to retrieved chunks, and without that rule the agent tends to quietly drop them.

How it works

At creation, the node builds a RagTool bound to your RAG_ENGINE and returns it along with a ruleset. Wire both into your agent - the TOOL gives it the capability, the RULES tell it how to behave. When the agent needs information, it calls the tool, which runs the query through your engine's stages and returns an answer with footnotes intact.

The inputs that matter

• rag_engine - required, and the heart of the node. You build this upstream: a Griptape RAG Engine node with retrieval and response modules connected. The tool is only as good as the engine you give it.
• description - required, default "Contains information about...". This is how the agent knows what this tool is for. The LLM reads it when deciding whether to call the tool, so write something concrete: "Contains product specs for the 2024 catalog" beats the placeholder every time.

Two outputs: TOOL (TOOL_LIST) and RULES (RULESET).

Install

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/griptape-ai/ComfyUI-Griptape

or ComfyUI Manager → search "Griptape" → install → restart.

Where people get burned

The most common failure is wiring this node without an engine - it's a required input, so it won't even build until you connect one. That's by design (the pack's simpler VectorStore tool exists for the "I just want retrieval" case). Second, people forget the RULES output. Connect it. That footnote rule is the difference between answers that cite their sources and answers that quietly lose them - the exact thing that makes RAG trustworthy.

Finally, a realistic note on effort: this node is the advanced path. If you're new to this pack, start with the Griptape Cloud KnowledgeBase tool (managed, zero setup) or the VectorStore tool (driver + description), and come back to the RAG engine once you understand retrieval stages and want control. When you do, the pack's RAG module nodes make it buildable without writing code - but buildable doesn't mean free, and there's real tuning in here.