Simple LLM: Vector Store

RAG - retrieval-augmented generation - is how you make an LLM answer from your documents instead of its training data, and it's all the rage. Simple LLM: Vector Store is this pack's entry point into that world: a node that creates a named store ready to hold documents. The name sounds like it's setting up serious semantic search. The reality is more modest, and knowing that up front saves you an hour of confusion.

It's part of the SimpleLLM family in ComfyUI-NS-Util (NakamuraShippo's utility pack, whose LLM section the README labels "Implementing the tests now"). Fair warning about what "vector" means here: the sibling node that adds documents (Simple LLM: Add Document) hashes each document with MD5 and calls that the "embedding," and the query node just grabs the first few documents in insertion order. That's a demo-grade RAG scaffold - a labeled box of text - not a real vector database with semantic similarity. If you need true semantic retrieval, you're better off with a dedicated RAG pack.

How it works

The node takes a name and returns a VECTOR_STORE object: an empty dict with a name, empty document/embedding/metadata slots, and a creation timestamp placeholder. It's a container, not a search engine. You then chain it into Simple LLM: Add Document to fill it (each add returns an updated store), and eventually into Simple LLM: RAG Query, which builds a "based on this context, answer the query" prompt you feed to a run node. The "retrieval" is insertion-order and top-k - fine for a handful of docs, misleading for hundreds.

The inputs that matter

• name - a string, default my_store. Just an identifier for the store; handy when you're juggling multiple document sets.

Output: a VECTOR_STORE - empty until you add documents.

Installing the pack

# ComfyUI Manager → Install via Git URL:
https://github.com/NakamuraShippo/ComfyUI-NS-Util

# or:
cd ComfyUI/custom_nodes
git clone https://github.com/NakamuraShippo/ComfyUI-NS-Util
pip install pyyaml watchdog

Restart ComfyUI. No model downloads - notably, no embedding model to fetch, which tells you exactly how much "vector" is going on here.

Common issues

The gap between the name and the behavior is the issue. Users come expecting semantic search and get ordered doc retrieval; if you're building a real RAG pipeline, budget for a proper vector-store node instead. Also, an empty store that gets queried returns a "No documents found" message - harmless, but easy to misread as a bug. And like the rest of the LLM family, this is new and under active development, so the "embeddings" scheme is likely to change. Good for learning RAG's shape inside ComfyUI; not yet a production tool.