Francisco

🚀 We've introduced llms.txt files for LangChain and LangGraph, supporting both Python & JavaScript! These help your IDEs & LLMs access the latest documentation. We also provide an MCP server to simplify IDE integration specifically for llms.txt files. Links below 👇

A Perfect Search Engine (8 min read) Isn't it strange that LLMs can solve graduate-level math questions, but today's search engines can't understand a 3 word phrase? For example, if you search "shirts without stripes" on google images, you get exclusively shirts that **do** have stripes. I started Exa to build a better search engine than Google, one that can actually understand you like an LLM does. As our company has grown and LLMs have advanced, so have our ambitions. We no longer just want search that's better than Google. We want search that's perfect. A perfect search engine would help the world more than people realize. And as far as I'm aware, Exa is the only organization on the planet building it. Let me tell you what we see. > The current search ecosystem Despite all the hype about AI search in recent years, search engines themselves haven't actually changed. Google search and Bing search are still very similar to what they were a decade ago. When you type a query like "shirts without stripes", they match the keywords in your query to all the documents on the web. The search fails here because "without stripes" requires understanding that goes beyond keyword matching. While search algorithms haven't changed, all the AI search hype has come from the introduction of AI summaries. Google AI overviews, SearchGPT, Perplexity -- each of these uses an old-school search engine like Google or Bing under the hood, and then uses an LLM to generate a summary. LLM summaries are great and save us time. But unfortunately the LLM is still limited by the quality of the underlying search engine. If Bing can't find something, then SearchGPT (which relies on Bing) can't either. And so today's "AI search" is mostly a time saver, it's not actually a better search engine that finds us things we couldn't before. I remember when I told my dad I was building a new search engine in 2021, he replied, "Isn't Google already good enough?". It's not good enough. Google's old school search algorithm works great for simple queries -- "Taylor Swift boyfriend" or "Walmart homepage". But it fails miserably the moment a query gets more complex -- "phds in the Bay area who've written about flying cars". If you ever go to Linkedin to find people, X to find interesting articles, or your friends to find good startups, then you are well aware that Google isn't good enough because you didn't choose Google for that information. But that's ridiculous because these are literally information searches over the web, and search engines in 2025 should be able to handle them correctly. It's not that Google hasn't *indexed* all this information -- Google has in fact indexed nearly all pages on the web (roughly a trillion pages). It's just that Google's algorithm was fundamentally not designed to handle complex searches over a trillion pages. But now there exist AI systems that can handle complex requests nearly perfectly. We should expect the same from our search engines. The world deserves a perfect search engine. > You really want a perfect search engine A perfect search engine is one that can find you any information you want no matter how complex your request. It's a database of humanity's collective knowledge, fully organized however you want in real time. That's hard to imagine because it doesn't yet exist, so here are some cool examples of what perfect search would be like: 1. Perfect search by idea Say you have an idea and you want to find similar ideas across the web. It's impossible to do that right now with traditional search engines. For example, I have this idea about flying cars that use roofs to recharge, and I want to find the hundreds of essays, tweets, youtube videos, and experts that discuss this idea and not some tangential idea. Google completely fails here, because this search requires real understanding. If I had a perfect search engine that could match content based on precise meaning, I'd instantly find everything I need. Perfect search by idea would have a huge impact across our lives. Here are some downstream effects: - Speeding up scientific progress -- "find all research papers that had a similar stem cell setup to mine" - A more informed society -- "give me all the counterarguments to my idea about immigration, found in British forums" - Identifying the frontier of human knowledge -- "has anyone ever proposed this moral framework before in the history of philosophy and if not what are the closest ones?" - Super fast idea gathering -- "recent tweets where people list ways OpenAI’s o3 will effect the future" 2. Perfect search for people We humans are constantly seeking out other people for friendship, collaboration, and community. But the best people search our advanced society has built is Linkedin -- ie basic keyword filters over linkedin profiles. The internet is so much richer than that. With perfect search, you should be able to find whoever you want. If I'm an undergrad who's studying AI alignment, I should be able to easily find "undergrads who took time off to study AI alignment and who have a blog". Try searching that on Google. Perfect web search would convert the messy web into a social network more powerful than Meta or X. Some major downstream effects: - Create communities on the fly -- "I'm creating a discord of best prompting practices. Find me anyone who's ever thought deeply about this" - Find collaborators in any city -- "Who in Berlin has written about new search algorithms?" - Fix recruiting -- "frontend engineers with design skills who recently left a big tech job and have startup experience" 3. Perfect multimodal search The web contains billions of videos, images, and songs, but because we can't search through them well, most of that value goes unseen. For example, I really like video clips where people make some sort of self-sacrifice for the world. Youtube search is horrendous at these searches. Perfect search would find you any media you want no matter how complex your search. Some major downstream effects: - Precise artistic exploration -- "images of shirts without stripes that have a cool mix of colors like a Jackson pollock" - Precise artistic exploration -- "find me hanz zimmer-like music with a fast beat similar to the song Supermarine" - Quickly gather supporting research -- "clips where scientists predict the effects of climate change from the 2010s" - Organize our visual world in beautiful ways -- "videos of baby's first words that aren't mamma or dadda" 4. Perfect control Google provides little control over search results. You type in a few keywords and hope it understands you. You never think to type a long sentence explaining exactly what you want because you know it won't work. Perfect search lets you add whatever filters you want, and it always works. If you're looking for people to hire, you'd be able to add modifiers like "who've worked at startups", complex conjunctions like "who know both rust and C++", and negatives like "who didn't study at these schools". The web would feel like a database you can filter arbitrarily however your mind wishes. Some major downstream effects: - Can finally seek out an unbiased understanding of the world -- first can search "analyses of X by liberal authors" and then "analyses of X by conservative authors" and then "analyses of X by conservative authors who used to be liberal", etc. - Search over any subset of the web -- "... and filter out all AI generated images" - Get your own personalized search engine -- "... and bias toward articles that a software engineer would like, someone who likes declarative sentences and hates clickbait" 5. Perfectly comprehensive We each walk around with an incomplete understanding of nearly everything. Whether you're looking for companies, blog posts, people, products, papers, etc, Google will give you a small sample of results, but you'll miss everything else that's out there. Never-ending FOMO. Perfect search fixes this -- if there are 387 entities in the world who match your criteria, you should be able to find all 387. Not 10, not 389, but 387. Exactly what you asked for. Perfect search would give us all a complete understanding of our world. Some major downstream effects: - True web analytics -- "what percentage of analyses of Elon's purchase of twitter from 2024 viewed it positively?" - Sort the web however you want -- "find the substack articles about climate change with the most comments" - Find the actual best of anything because you know it's comprehensive -- "cheapest electric skateboard that can go faster than 22mph and has good reviews" - Realtime trends -- "What is the breakdown of what people from South America are commenting about US politics" > Agents + perfect search Perfect search is most powerful when paired not with a human, but with an agentic LLM (which is coming in 2025). The human chats with the LLM and the LLM agentically searches in the background, potentially hundreds of times. As a final example, imagine I have an idea for how flying cars can leverage city infrastructure. I ask my LLM to generate a report on everything I can do to make this work. The LLM first searches for all similar ideas across the web, in posts, tweets, research papers, news articles, videos, etc. For each idea, the LLM finds the best counterarguments on the web if they exist. The LLM then creates a fully informed step by step plan based on all those ideas. Then for each step of the plan, the LLM finds people in my current city who can be helpful to implement the step, from suppliers to independent researchers. The LLM gathers the contact info of the ones with publicly available info and creates a personalized message for each. With an LLM + perfect search, it takes a few minutes to gather a comprehensive action plan. With an LLM + traditional search, it would never fully work -- I'd forever only know a small percentage of all the relevant ideas and people out there, and my flying car idea would never take off. > Why perfect search matters I've only listed a small subset of all the downstream effects of perfect search. Similar to LLMs, it's difficult to predict all the ways perfect search would be used and integrated into our tools. I'd go as far as to say that the future of humanity depends on getting this right. Because the information we consume has immense power over us. It determines what we know, which determines how we think, which determines how we act. Right now the world's information is a complete mess. There's a firehose of content pouring onto the world each day, without any sort of organization. Is it any wonder that people's perspectives are all over the place? It's hard to know what is true, what is radical, what is reasonable, who thinks what, why they think it, where did those ideas originate, etc. Despite all the advances in AI, our society’s main information tools are still Google, social media, and the mainstream media. None of these are capable of nor devoted to providing a comprehensive, unbiased, high quality understanding of the world. This is a dangerous situation for a nuclear armed society rapidly approaching AGI. We need to fix our information ecosystem, so that we can navigate the next decade as a well-informed species. Perfect search is by definition a solution to our information ecosystem because it gives everyone full control over the world's information -- our information -- so that it's digestible, legible, useful, actionable. Perfect search is civilizational infrastructure that needs to exist and no one else is building it. That's why Exa is building it. > Why we're going to do it So why has no one built a perfect search engine yet? Three reasons -- money, technology, and insanity. To build perfect search, you need an organization with the right financial incentives. Google makes 200 billion dollars a year from search ads. Perfect search wouldn’t make Google more ad revenue – it could even make them less. Exa makes money from API usage and subscriptions. We are strongly incentivized to improve search until it’s perfect. To build perfect search, you need to redesign the search algorithm using novel neural methods, not keyword methods. Traditional search engines were designed two decades ago when computers could not think. Neural search engines are more chaotic and unpredictable, but in time they will win over traditional ones. The big players won’t build neural search engines because their infrastructure and products are built around an old paradigm. And for new players it’s quite hard to develop the novel ML architectures and web-scale infrastructure that’s required. At Exa, we have the freedom, experience, and now resources to build search algorithms from scratch the right way. To build perfect search, you need to be a bit insane. Since we started the company in 2021, pretty much everyone thought what we were doing was insane. “What’s wrong with search?” “Isn’t Google good enough?” You need a group of people who refuse to accept the way things are and work for years until an abstract vision is made real, even if no one else gets it. That’s what OpenAI did with intelligence, and it’s what we intend to do with knowledge.