Docq.AI retweeted
Chunking techniques for RAG:
Fixed-length: Simple but rigid.
Sentence-based: Natural but variable size.
Paragraph-based: Logical but inconsistent lengths.
Recursive chunking: Flexible but uses different logic.
Semantic: Context-aware but computationally expensive.
Sliding window: Flexible overlap but potential redundancy.
Document-based: Preserves structure but may be too large.
Experiment with different methods. Find what works for your data.
Docq.AI retweeted
The improved version of the @docqai Chrome extension is coming along nicely.
UI is powered by @chakra_ui.
Creating that custom component for Space selection was 10x easier, faster, and more enjoyable to dev with Chakra (+ Copilot).
ALT screenshot of the Docq Chrome extension
Docq.AI retweeted
This OpenAI structured outputs feature, released the other day, is a big deal.
It might not sound cool/sexy but it's important. Anybody building real AI products knows that _reliably_ binding LLM output to the UI and other parts of the software system is hard. >>
Docq.AI retweeted
Boom I'm organising another Space tomorrow!
We'll cover the latest AI/LLM news and talk about what each of us is working on.
Join if interested!
xitter.cf/i/spaces/1yoJMyNvXywJQ
Finetuning models shouldn't be your opening move!
medium.com/@docqai/leveragin…
Docq.AI retweeted
StreamPot is now working 🫖
Plz try it. Plz roast it. Plz use it for all your video processing needs.
Docq.AI retweeted
Those building RAG solutions: have you tried HyDE query rewrites? what's your experience of its performance? if not great, what's working better?
HyDE (Hypothetical Document Embeddings) paper: arxiv.org/abs/2212.10496
Docq.AI retweeted
Rewrote the core RAG logic in @docqai today to use declarative query pipelines in Llama Index. We were using the RAG fusion retriever + vector and BM25 search. Everything was wrapped into an inflexible box. So the work involved recreating that from composable pieces. >>
Docq.AI retweeted
In April, there were four major open LLM releases: Mixtral, Llama 3, Phi-3, and OpenELM.
In his latest article, @rasbt summarizes and reviews the latest LLM releases and covers everything from performance to available training information.
Read more: magazine.sebastianraschka.co…
Docq.AI retweeted
I’m starting to suspect that the target audience for @chatprd is non PMs and I think that PMs should think hard about what that means for the future of the role.
The biggest impact that ChatPRD is having at @wearefitxr is reducing the mental aversion non-PMs have to writing a PRD. It massively speeds up the process but most importantly it makes it feel easier / less scary.