Midnight 🌙
@midnightai.app
Search you can talk to 🌙
Describe what you're looking for in a normal sentence - no keywords, no operators, no quotes around phrases.
Private beta.
Describe what you're looking for in a normal sentence - no keywords, no operators, no quotes around phrases.
Private beta.
Pinned
Midnight 🌙
@midnightai.app
· Jul 30
Hi! I'm building a search engine called Midnight.
premise: you describe what you want in a normal sentence, it finds it. no keyword archaeology, no quotes-around-phrases, no site: operators.
Some features that aren't working in the beta test yet, but will be available soon
Site: midnightai.app
premise: you describe what you want in a normal sentence, it finds it. no keyword archaeology, no quotes-around-phrases, no site: operators.
Some features that aren't working in the beta test yet, but will be available soon
Site: midnightai.app
Watched someone type the same query three times today. Different words, identical meaning, zero tags in any of them.
That's the real input. Not the clean query in your test suite.
Eval sets built at a desk don't survive contact with how people actually type.
#midnight #semanticsearch #search
That's the real input. Not the clean query in your test suite.
Eval sets built at a desk don't survive contact with how people actually type.
#midnight #semanticsearch #search
September 24, 2026 at 7:03 AM
Watched someone type the same query three times today. Different words, identical meaning, zero tags in any of them.
That's the real input. Not the clean query in your test suite.
Eval sets built at a desk don't survive contact with how people actually type.
#midnight #semanticsearch #search
That's the real input. Not the clean query in your test suite.
Eval sets built at a desk don't survive contact with how people actually type.
#midnight #semanticsearch #search
How it works, short version:
Everything gets turned into a vector - numbers describing what's in it,
not what it's labeled.
Your sentence becomes a vector too. Then it's just distance.
Drop an image and you skip the words entirely. Same space, same math.
Everything gets turned into a vector - numbers describing what's in it,
not what it's labeled.
Your sentence becomes a vector too. Then it's just distance.
Drop an image and you skip the words entirely. Same space, same math.
September 22, 2026 at 7:51 PM
How it works, short version:
Everything gets turned into a vector - numbers describing what's in it,
not what it's labeled.
Your sentence becomes a vector too. Then it's just distance.
Drop an image and you skip the words entirely. Same space, same math.
Everything gets turned into a vector - numbers describing what's in it,
not what it's labeled.
Your sentence becomes a vector too. Then it's just distance.
Drop an image and you skip the words entirely. Same space, same math.
Tags describe the file. They don't describe the thing.
Every item you search is indexed by whatever someone typed in a hurry,
years ago, optimizing for something that wasn't your question.
Embeddings index what's actually there. That's the whole idea.
Every item you search is indexed by whatever someone typed in a hurry,
years ago, optimizing for something that wasn't your question.
Embeddings index what's actually there. That's the whole idea.
September 20, 2026 at 8:14 PM
Tags describe the file. They don't describe the thing.
Every item you search is indexed by whatever someone typed in a hurry,
years ago, optimizing for something that wasn't your question.
Embeddings index what's actually there. That's the whole idea.
Every item you search is indexed by whatever someone typed in a hurry,
years ago, optimizing for something that wasn't your question.
Embeddings index what's actually there. That's the whole idea.
Everyone asks which vector DB to pick.
We rewrote our chunking three times. That moved recall more than every model swap combined.
The database isn't the bottleneck. The split is 🌙
We rewrote our chunking three times. That moved recall more than every model swap combined.
The database isn't the bottleneck. The split is 🌙
September 19, 2026 at 2:07 PM
Everyone asks which vector DB to pick.
We rewrote our chunking three times. That moved recall more than every model swap combined.
The database isn't the bottleneck. The split is 🌙
We rewrote our chunking three times. That moved recall more than every model swap combined.
The database isn't the bottleneck. The split is 🌙
Two ways to search, no keywords in either one.
Type a sentence the way you'd say it out loud. Or drop an image and let
the embeddings do the work - ranked by cosine similarity, scored to the percent.
Private beta. Building it in the open.
Type a sentence the way you'd say it out loud. Or drop an image and let
the embeddings do the work - ranked by cosine similarity, scored to the percent.
Private beta. Building it in the open.
September 13, 2026 at 6:32 PM
Two ways to search, no keywords in either one.
Type a sentence the way you'd say it out loud. Or drop an image and let
the embeddings do the work - ranked by cosine similarity, scored to the percent.
Private beta. Building it in the open.
Type a sentence the way you'd say it out loud. Or drop an image and let
the embeddings do the work - ranked by cosine similarity, scored to the percent.
Private beta. Building it in the open.
Hi! I'm building a search engine called Midnight.
premise: you describe what you want in a normal sentence, it finds it. no keyword archaeology, no quotes-around-phrases, no site: operators.
Some features that aren't working in the beta test yet, but will be available soon
Site: midnightai.app
premise: you describe what you want in a normal sentence, it finds it. no keyword archaeology, no quotes-around-phrases, no site: operators.
Some features that aren't working in the beta test yet, but will be available soon
Site: midnightai.app
July 30, 2026 at 9:05 AM
Hi! I'm building a search engine called Midnight.
premise: you describe what you want in a normal sentence, it finds it. no keyword archaeology, no quotes-around-phrases, no site: operators.
Some features that aren't working in the beta test yet, but will be available soon
Site: midnightai.app
premise: you describe what you want in a normal sentence, it finds it. no keyword archaeology, no quotes-around-phrases, no site: operators.
Some features that aren't working in the beta test yet, but will be available soon
Site: midnightai.app