[Backend] Use paragraphs instead of words in chromadb - Githubissues

apirrone / Memento

Memento is a Python app that records everything you do on your computer and lets you go back in time, search, and chat with a LLM (Large Language Model) to find back information about what you did.

MIT License

484 stars 39 forks source link

[Backend] Use paragraphs instead of words in chromadb #6

Closed apirrone closed 11 months ago

apirrone commented 11 months ago

Adding a lot of independant words to chromadb takes a long time ans is inefficient in terms of semantic search
Try to merge the bboxes to make paragraphs, the ocr can run efficiently on crops of images
then when querying, get the paragraph and re run ocr to get the precise word ?

apirrone commented 11 months ago

It is WAY faster to embed paragraphs than individual words Capture d’écran du 2023-08-04 09-35-19

apirrone commented 11 months ago

Done, works a lot better

apirrone commented 11 months ago

Actually, maybe we should retrieve by frame with chromadb, then do a ocr search for each retrieved frame

apirrone commented 11 months ago

Actually, maybe we should retrieve by frame with chromadb, then do a ocr search for each retrieved frame

This is what is done right now