Metaphor is a search engine designed from scratch using AI. Our goal is to find what your smartest friend could find you in a week, but instantaneously. By discovering new ways to search the internet, we'll rediscover the internet.
Imagine being able to search the entire repository of humanity's knowledge (the internet) with queries of any complexity (think multiple sentences!) and find exactly the results that match your query. AI has gotten so good that this is now possible. Metaphor's current model is a step in that direction.
Metaphor's AI model is trained to do link prediction. This means that given some text prompt, it tries to predict the link that would most likely follow that prompt. Because of the way our current model was trained, the best prompts resemble how someone on the internet might refer to a link. (We're working on fixing this model quirk, but for now these guidelines will help make your search experience way better!)
Suppose you're interested in motorcycle riding and want to find the personal sites of interesting motorcycle people.
You might think something like “Motorcycle riding personal page” or maybe something like “People interested in motorcycles” would be good prompts.
But if you try searching for either of these, you’ll see that the results aren’t very dialed in. The results are poor because it's not how someone would describe a personal site they’re about to share.
Here's a better version of the prompt:
"This person would be great to talk to about motorcycles (personal site here:"
This prompt resembles the way someone on the internet might actually talk about a link they found. You can imagine this prompt on a particularly helpful subreddit or in the middle of a good blog post.
Note how near the end of the prompt we specify what kind of website we’d like Metaphor to return – “personal site”. If, for example, we're searching for a blog post or arxiv link we should specify that by adding “blog post here:” or “arxiv paper:” to the end of the prompt.
Suppose you were interested in finding the best math blogs on the internet.
Here's a bad prompt: “What are the best math blogs on the internet?”
This is not a very natural way to precede a link.
Instead, someone searching for great math blogs might already have a great math blog in mind that they wish to find content similar to. We could try a prompt like
“Besides Terry Tao’s blog, this is my favorite math blog:”
This gives much higher quality results.
When possible, try to avoid keyword searches
The results tend to be less robust and useful, because they look very different from the type of ways people might talk about links on the internet. Instead of typing “Jeopardy archive” try something like “Here is the Jeopardy archive:”
Rephrase questions to look more like answers.
Questions tend to be bad prompts because people sharing links on the internet don’t usually precede links with questions. If you add "I found a good answer here:" after the question the results are often better.
For example, instead of searching for “What’s the best way to get started with cooking?”, try rephrase the question to resemble an answer;
“This is the best tutorial on how to get started with cooking:”
Use modifiers to control the type of results returned.
If you’re looking for a particular kind or style of result (a goodreads link, or a funny post) try specifying that in the prompt.
Pay attention to the punctuation at the end of prompts.
You’ll notice many of the example prompts above end in colons “:” because that mimics how someone would share a link.
Our current models are just the beginning. Search systems of the future will feel like extensions of your mind – capable of rich interactions across multiple modalities and multiple reasoning steps, weaving together unique insights from the sum total of all knowledge.
If you find this vision compelling – and if the future of search is one you’d like to shape – come join us on the journey!
Let us know what you think at [email protected]!