I'm wondering if we can integrate AI into search tools in someway. It would be very nice to have more information about files than just the filename. What if we could run AI in the background (using LMStudio or others to use local or remote) where it analyzes our files to get information about them then rather than just searching by filename we can essentially search by embedding information. This wouldn't be the same as contents search. It would be a more generalized search. E.g., "Show me all files related to microscopy" and it would then query such files. Furthermore, it could collect information online about certain common files, handle tools and file information, etc. We could then also interact with search everything in a more organic way such as "Show me all the files that were created by the system yesterday between 2pm and 3pm".
Who knows what else it could do. Contents can be summarized when available, we can add content information for binary files that are just data, images could be analyze to be able to deal with large image sets that properly named, find duplicates more easily, etc.
AI integration
-
rezolva
- Posts: 1
- Joined: Tue Jun 09, 2026 9:40 am
Re: AI integration
Disclosure: I'm the developer of https://TheFile.Ninja, a file manager built around Everything, so this is something I've been working on directly.
I'd split the idea in two layers:
1. Natural language → query. In TheFile.Ninja you type something like "files created yesterday between 2pm and 3pm", and an LLM translates it into Everything search syntax. It returns 2–4 alternative queries with short descriptions, you pick one, and Everything does the actual search, so it stays instant. It's bring-your-own-key. What made it work was giving the model a compact reference of Everything's syntax that only covers what affects the query itself.
2. What content search can't do. Everything can already index file contents (source code, text, etc.) if you enable it, and TheFile.Ninja integrates that. That covers a lot of "find the file that mentions X". The gap is where the content isn't searchable text: images, binary data, and abstract queries like "everything related to microscopy". That's where an AI layer adds something new. I generate AI descriptions for images with an lua automation and show them in a Description column. Embeddings for everything on a drive would be a much heavier step (extraction, vector index, incremental updates), so I'd want it opt-in, per folder, and local-first. Local models are on my roadmap for that reason (Next version have a experimental version of this feature).
Happy to share what I learned about prompting for EQL.
I'd split the idea in two layers:
1. Natural language → query. In TheFile.Ninja you type something like "files created yesterday between 2pm and 3pm", and an LLM translates it into Everything search syntax. It returns 2–4 alternative queries with short descriptions, you pick one, and Everything does the actual search, so it stays instant. It's bring-your-own-key. What made it work was giving the model a compact reference of Everything's syntax that only covers what affects the query itself.
2. What content search can't do. Everything can already index file contents (source code, text, etc.) if you enable it, and TheFile.Ninja integrates that. That covers a lot of "find the file that mentions X". The gap is where the content isn't searchable text: images, binary data, and abstract queries like "everything related to microscopy". That's where an AI layer adds something new. I generate AI descriptions for images with an lua automation and show them in a Description column. Embeddings for everything on a drive would be a much heavier step (extraction, vector index, incremental updates), so I'd want it opt-in, per folder, and local-first. Local models are on my roadmap for that reason (Next version have a experimental version of this feature).
Happy to share what I learned about prompting for EQL.
You do not have the required permissions to view the files attached to this post.
Last edited by rezolva on Thu Oct 08, 2026 8:11 am, edited 2 times in total.