Griptape Rerank Driver: Local

If you want to make your RAG results sharper and you don't want to pay Cohere per query, this is the node. The Griptape Rerank Driver: Local does the second-pass ranking pass on your own machine, and it is gloriously boring: no inputs, no API key, no network calls at runtime.

That's the whole pitch. It builds a LocalRerankDriver, which runs a small reranking model locally - a cross-encoder that scores how well each retrieved chunk actually answers the query, instead of just measuring vector distance. The first time it runs it downloads the model from Hugging Face, so make sure you have a connection (and a few hundred MB free) on that first run. After that it's local.

You'd reach for this over the Cohere variant when: you're already running everything local with Ollama, you're iterating on a RAG pipeline and don't want per-query costs while you tune, or you just value not having yet another key in your settings. The tradeoff is speed and raw quality - Cohere's hosted rerankers are stronger and faster on big result sets, and a local cross-encoder adds CPU time per query. For a personal pipeline with a few results to rank, you'll never notice.

How it works

Drop it into a Griptape Rag Engine node's rerank_module input (or wire it to a RAG tool's rerank stage). The driver re-scores the retrieval stage's candidates before the response stage writes an answer. The node has zero configurable inputs - you connect it and it's done - and emits a single DRIVER of type RERANK_DRIVER.

Installing

Same pack as everything else here. ComfyUI Manager → search "Griptape" → Install, or:

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

Restart ComfyUI. The pack installs griptape[all] and python-dotenv; the first install is heavy, and the first run of this node adds a model download on top.

Common gotchas

The one real surprise is that first-run model download - if it fails with a network error, check your Hugging Face connectivity rather than your ComfyUI setup. Also remember this node only helps if something actually uses the driver; an unconnected rerank module means it never runs.

Then the pack-wide usuals: torch version conflicts on Nvidia (reinstall with --extra-index-url https://download.pytorch.org/whl/cu121) and stale-griptape ImportErrors (fix: python -m pip install griptape -U).