Spencer Miskoviak

[Asktro @asktro_]

Asktro now provides a plugin and UI support for @docusaurus 🦖

In just a few lines of code, you can add document search and an AI assistant tailored to your documentation.

See a demo: https://t.co/3zjqxBiAiR

Get started: https://t.co/kQrL3eYZNi

[Spencer Miskoviak @spencerskovy]

LLM Markdown (a demo to render rich responses from LLMs) now supports LaTeX as well. 🧮

Give it a try: https://t.co/nTV9QrgTUD

[Asktro @asktro_]

Want to see a production-like demo of how Asktro works? 🔎

Checkout this fork of the Nextra docs with Asktro search enabled to compare with the default search results. 📊

Asktro Nextra docs demo: https://t.co/rYlnsxrknP
Nextra docs: https://t.co/BcbQsrjU5O

[Spencer Miskoviak @spencerskovy]

It took some experimentation to build the @asktro_ UI component package with tsup and Tailwind, and to output the CSS in a way to avoid class name collisions, avoid global resets, and support dark mode.

These were the steps to get it working:

https://t.co/0NsV5iWz7M

[Spencer Miskoviak @spencerskovy]

Spent longer than I should've animating the @asktro_ logo but @khaicp's design and hover animation concept was too good not to build 👨‍🚀🌟

[Spencer Miskoviak @spencerskovy]

@supabase @qdrant_engine The search UI is built with @tailwindcss, cmdk, @vercel AI, @_phosphoricons, and more

https://t.co/WHmMs24N6D

[Spencer Miskoviak @spencerskovy]

🔍 Exploring a rich search experience for static documentation

💽 A webpack plugin extracts the Nextra docs which are embedded/stored with @qdrant_engine

✨ The custom search component is built with cmdk from @pacocoursey

🤖 Next up: use an LLM to answer directly based on docs

[Spencer Miskoviak @spencerskovy]

Excited to share a small project I've been working on inspired by the rich search experience @supabase built for their docs.

Asktro is a Nextra plugin that extracts, embeds, and persists the docs in @qdrant_engine and surfaces the search and AI-assistant in static docs.

[Asktro @asktro_]

🚀 Asktro is live for early access!

👨‍🚀 Asktro brings a dynamic, AI-enabled search experience to static documentation (currently, Nextra).

🔎 Powered by the latest embedded text similarity search and large language models.

[Erfan @ErfanEbrahimnia]

Learn how to use Vercel’s AI SDK with Next.js to build an AI app that understands markup languages for formatting and rendering diagrams in MermaidJS

https://t.co/xmPheUlBeT

[Erfan @ErfanEbrahimnia]

🔥 The 5 hottest Next.js posts of this week – Next.js Weekly #15

◆ Demystifying RSCs
◆ Rendering rich responses from LLMs
◆ Visual Guide to Prefetching
◆ Introducing: Kuma UI 🐻‍❄️
◆ Invisible Details of Interaction Design

in detail 👇

[Spencer Miskoviak @spencerskovy]

🔍 Exploring a rich search experience for static documentation

💽 A webpack plugin extracts the Nextra docs which are embedded/stored with @qdrant_engine

✨ The custom search component is built with cmdk from @pacocoursey

🤖 Next up: use an LLM to answer directly based on docs

[Spencer Miskoviak @spencerskovy]

An overview for rendering rich responses 🎨 (markdown and diagrams) from OpenAI 🤖 with @nextjs, @vercel ai SDK, @unifiedjs, and @mermaidjs_.

📝 https://t.co/goKymFgidX

[Spencer Miskoviak @spencerskovy]

LLMs like GPT are capable of producing responses with popular formatting like Markdown, and Mermaid diagrams. 📊

LLM Markdown demonstrates how to leverage this with @vercel AI, @unifiedjs, and @mermaidjs_ to render rich-text responses. 🎨

🖥 Demo: https://t.co/nTV9QrgTUD

[Spencer Miskoviak @spencerskovy]

@remcohaszing @vercel @unifiedjs @mermaidjs_ Nice, that would be useful!

A challenge is Mermaidjs isn't as forgiving at parsing invalid diagrams. Markdown is more forgiving rendering invalid while streaming. This is why the demo required a click to render the diagrams. I imagine this plugin would run into the same issue?

[Spencer Miskoviak @spencerskovy]

LLMs like GPT are capable of producing responses with popular formatting like Markdown, and Mermaid diagrams. 📊

LLM Markdown demonstrates how to leverage this with @vercel AI, @unifiedjs, and @mermaidjs_ to render rich-text responses. 🎨

🖥 Demo: https://t.co/nTV9QrgTUD

[Spencer Miskoviak @spencerskovy]

🤖 A system prompt is used to encourage this formatting in the response. The response is then parsed and rendered with popular JavaScript libraries to turn the text-only formatting into richer responses.

📝 Source: https://t.co/7zuX7NThNx

[Spencer Miskoviak @spencerskovy]

Lots of ideas on how to improve this with AST/LLM-based chunking, hierarchical indexing, metadata generation (summaries), query expansion, etc.

Excited to see what this tool can unlock!

[Spencer Miskoviak @spencerskovy]

It’s been fun building this, learning new techniques, and watching how fast the tooling is evolving in the space 🤯

At a high-level, how we built this…👇

[Naren Manoharan @narenmano]

We built a tool to help you not copy paste code into ChatGPT! 🤖

https://t.co/K0rOCZbcpL

[Spencer Miskoviak @spencerskovy]

- Chunk each file into smaller pieces
- Vectorize each chunk/metadata with embeddings and persisted in a vector store
- The query is also vectorized and compared to chunks for context
- Construct a prompt with the original query and relevant context
- Send the prompt to an LLM

[Fabien Devos @Fabien_Devos]

New blog post! 📰 How to ship higher quality apps faster by adopting mobile release trains: https://t.co/9cWUum8wEM

And by the way, @wolfia_app can automate your mobile app releases 🚀, including setting up a CI for free! Ping us at [email protected] or DM me to learn more.

[Wealthfront Eng @wltheng]

We've fully migrated the Wealthfront web app to TypeScript and are excited to share our journey, and how we plan to continue improving the type safety in our codebase! https://t.co/cMeECkwYTj

[Spencer Miskoviak @spencerskovy]

@bjreath @scottringwelski @OkDrewgle Phew.

[Spencer Miskoviak @spencerskovy]

Trying to build an MDX powered blog with Nextjs? It can be challenging to use the Nextjs Image component with mdx-bundler to get all the experience and performance image optimizations.

Here’s how I solved it for my blog: https://t.co/2pcuQ6VYj7