X Bookmarks — 2024 KW10: Long Context RAG and Canvas UIs

March 7, 2024

|bookmarks

by Florian Narr

X Bookmarks — 2024 KW10: Long Context RAG and Canvas UIs

@LangChain — RAPTOR: long context RAG without chunking

Building long context RAG from scratch with RAPTOR + Claude3 (Video)

The rise of long context LLMs + embeddings will change RAG design. Instead of splitting docs + indexing doc chunks, it's feasible to index full docs.

But, there is a challenge to flexibly answer lower-level questions from these docs...

Interesting shift in RAG thinking. The standard playbook is chunk → embed → retrieve, but as context windows get longer and cheaper, indexing full documents becomes viable. RAPTOR takes it further by building a tree of summaries at different levels of abstraction, so you can answer both fine-grained and high-level questions from the same index — without losing signal by cutting the doc into arbitrary pieces. Worth watching if you're building anything retrieval-heavy.


@Mike_Andreuzza — uvcanvas, open source React canvas components

An open source React.js component library for beautifully shaded canvas, brought to you by Latent Cat.

  • uvcanvas .com

Smart timing to release this. Canvas-based UI components have always been annoying to drop into React apps cleanly — you end up wrapping imperative canvas code in useEffect and hoping for the best. A component library that handles the lifecycle management and ships a clean API is genuinely useful. The "beautifully shaded" part I'll believe when I see it in production, but the abstraction itself is the right call.


@LangChain — SalesGPT, an AI-driven sales agent

💸SalesGPT: Elevating Sales Conversations with LangChain

SalesGPT is an awesome open-source repo that creates an AI-driven sales assistant

This blog by @andysingal deep dives into how this agent is created and how to use it

Honestly saved this more for the architecture pattern than the sales angle. Building a stateful conversational agent that tracks conversation stage, decides when to push for a close, and adapts its pitch is a non-trivial orchestration problem. The repo is a decent reference for how to structure a multi-step agent with memory and conditional branching in LangChain — strip out the sales persona and the underlying pattern applies to a lot of things.


@stanley_human — Awwwards-winning portfolio

My old portfolio site. Snatched awwwards with this bad boy

Not software engineering, but this one's in here for the craft. Portfolio sites that win Awwwards are a useful benchmark for what's possible on the web without reaching for a design tool — just code, WebGL, animation timing. Useful to look at when you're about to ship something you think is "polished" and need to recalibrate.