January 28, 2025

New in Lex: Web search! (great for fact checking) and lots more

Hey! Nathan here, founder of Lex (https://lex.page), the modern writing platform with AI tools built in.

This week is a big one! We shipped:

1. Web search (using Perplexity) for Lex Pro users

2. DeepSeek R1 support

3. The Lex Writers Series. Our first episode features Carl Richards, bestselling author, NYTimes columnist, and Lex power user.

4. A new Lex information hub

5. Some meaningful performance improvements, and more

Let’s get into it:

---

1. Web search

We love Perplexity. It’s amazing at quickly searching the web and having AI summarize the results. This is an amazing tool for fact checking and ensuring no hallucinations, and especially helpful with niche or timely topics.

Until now there wasn’t an easy way to bring Perplexity into our writing workflows with Lex. But today we just shipped a massive leap forward.

Here’s how it works:

When you're writing and you want to look something up, there are three distinct steps:

1. Understand what I actually want and translate it into a search query

2. Perform the search and interpret the results

3. Integrate the results into my document

Our new web search feature performs each step in the most intelligent way possible:

1. We look at your document, the current chat thread, and anything else you'd like to add to the context window (for example, context tags) and use that information to understand what you want, and generate a web search query for you.

2. Then, we use Perplexity to perform the search and summarize the results.

3. Finally, we pass the web results—along with your document, chat history, context tags, etc—to the LLM of your choice for overall analysis and integration into your document, including specific proposed edits.

Web search is available today for Lex Pro subscribers! Just hit the ‘web search’ toggle in Ask Lex to test it out!

2. DeepSeek R1

There’s a hot new model in town: DeepSeek R1 is the first open source “reasoning model” to show comparable performance with OpenAI’s o1 on several benchmarks.

Start today, DeepSeek R1 is available in Lex Pro! This model is built for analytical thinking, problem-solving, and handling complex topics with nuance. When you use it in Lex, it’ll outline it’s thinking before responding.

We’ve had a lot of fun testing it out, and can’t wait to hear how the Lex community puts it to use!

PS—The privacy-conscious among you may be curious which provider we use to serve open-source models like DeepSeek R1. The answer is Together.ai, an American company that has an excellent privacy policy and does not retain or train on your data.

PPS—Fun fact, I fact checked the above statement about Together.ai using our new Web search feature and DeepSeek, and it worked beautifully! (I verified manually, too.)

3. Lex Writer Series - Episode 1

In the first episode of our new video series, we meet with Carl Richards (former New York Times columnist and bestselling author of The Behavior Gap and The One-Page Financial Plan) to explore one of our favorite Lex use-cases: using your past writing to teach AI to be an editor that brings out the best of your personal writing style.

Give it a watch and tell us what you think!

4. All your Lex questions answered in one place

We've just launched Lex Docs – a collection of essays (including the 2024 Annual Letter we released last week), how-to guides, info about Pro, and policies. Whether you're new to Lex or a power user looking to make sure you’re getting the most out of a feature, you'll find what you need here.

Of course, if you still have questions, we’re always happy to help at hello@lex.page!

5. Performance + miscellany

A couple months ago, we mentioned that we were investigating some issues around performance and connecting with our AI providers. We've made significant progress on two fronts:

1. The calls were coming from outside the house: We discovered that many of the issues were actually coming from our AI partners' outages, not Lex itself. Going forward, you'll get a clear notification when there are AI provider issues so you can switch to a different AI model and keep working without interruption.

2. We still improved performance: This exercise exposed a few ways we could make Lex faster and more reliable across the board. We shipped those improvements too and especially on longer docs with a lot of edit history you may have noticed significant improvements.

3. We shipped some other nice things:

   1. We improved the prompt that generates chat thread titles to be less repetitive and more descriptive

   2. We tweaked the system prompt in Ask Lex to be a bit less eager to suggest specific edits, although we still have more work to do here.

   3. We fixed a bug with Context Tags that prevented users from deleting attachments

   4. We added keyboard shortcut hints for adding a comment (hint: it’s the same shortcut as google docs)

We have more work to do, but we’re happy with the progress so far!

---

That's it for this week! Thank you for your time and support – it means the world to us.

—Nathan