Edit Embeddings Database

Edit Embeddings Database is EditResults' bigger sibling, applied one level earlier in the pipeline. Where EditResults filters the results of a search, EditDB filters the database itself before you ever search it. Same three methods, same wildcard matching, but the input is a LoadDB object and the output is a new, edited database.

Why would you bother? Three reasons, in rough order of usefulness:

• Shrink before searching. A 100,000-image database is fine to search, but if you only care about results from *beach* folders, pruning the DB down first makes every downstream step faster and cleaner - and scores stay relative to what's left.
• Re-point a moved folder. Same "replace" trick as EditResults: images moved, paths now 404, so rewrite the old path prefix to the new location in the DB without regenerating any embeddings.
• Segment a shared DB. One big collection, multiple focused searches against different subsets. EditDB gives each branch its own trimmed copy.

How it works

The inputs mirror EditResults: img_db (the LoadDB output), method (exclude to drop matches, filter to keep only matches, replace to rewrite matched path text with replace_text), and edit_text for the wildcard pattern (* matches any run of characters, fnmatch style). The node walks every entry in the loaded DB, applies the rule to each stored filename, and hands back a fresh IMG_DB output you can wire into Image Searcher or chain into another EditDB.

One important detail: the edit is in-memory only. Your new_img_db.json on disk is untouched; what you get back is a filtered view living in the current graph. If you want the trimmed database persisted, regenerate it (or write your own export) - this node won't do it for you.

Installing it

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/MoonMoon82/ClipVision_Tools
cd ClipVision_Tools
python -m pip install -r requirements.txt

Or via ComfyUI Manager ("ClipVision_Tools"). Restart after. Deps are orjson and pillow-heif only.

Gotchas

• It edits paths, not content. The embeddings stay whatever they were. Filtering out a folder doesn't improve how the remaining images are represented - it just removes them from consideration.
• Keep the model rule in mind: if you're chaining multiple LoadDBs or edited DBs together, they all have to come from the same CLIP vision model or the search is meaningless.
• The replace pattern strips the * you put in the match: edit_text oldpath* with replace_text /newpath/ rewrites the matched prefix. Check one result before trusting the whole DB.
• Wildcards, not regex. * and ? only. Don't bring your regex muscle memory.

It's a small, unglamorous node, but it's the difference between "search this 100k collection" and "search the 2k images that are actually relevant to this job." For anyone building a reference-library workflow, that's the node you'll quietly reach for every day.