Fangrui Liu

Data Scientist / Researcher, UBC EE MASc Metric Learning / Computer Vision

[Fangrui Liu @mpsk_liu]

Takes less effort than I thought. Thanks to @langchain ’s vector sql chain and self-querying retrievers!

Now you can ask with 36 million paragraphs under 5 million titles from Wikipedia! No more browsing or clicking, just write down your questions and get the answer!

[MyScale @MyScaleDB]

🥳Introducing an excellent update on ChatData! Wikipedia is now available as a knowledge base!

😉Come unlock the true potential of information retrieval with ChatData and discover a world of knowledge at your fingertips.
https://t.co/cnA4mEVBns

#selfquery #text2sql #vectorsql #vectordatabase

[MyScale @MyScaleDB]

🎉 Exciting news! MyScale custom GPTs is now available on the @OpenAI GPT store! 🚀Elevate your chat experience with enriched knowledge from ArXiv and Wikipedia via the @MyScaleDB.
🤗Come and explore now: https://t.co/DS0IyrimPj
#MyScaleGPT #vectordatabase

[Fangrui Liu @mpsk_liu]

Thanks for the opportunity @matthew_d_white and personally I think RAG is the future to scalable LLMs, which can make use of millions of external knowledge.

[MyScale @MyScaleDB]

📽️ Meetup recording of "Evaluating and Optimizing RAG Performance" is now available!
🚀 @MyScaleDB data scientist Fangrui Liu @mpsk_liu shared invaluable insights that can benefit your #RAG evaluation knowledge. Don't miss out: https://t.co/oN4j9EkpMt

[MyScale @MyScaleDB]

😖Evaluating RAG systems can be difficult.
🤗Join @mpsk_liu's talk introducing:
✅an open-sourced end-to-end benchmark framework on RAG system developed by #MyScale
💡some insights and interesting findings on performance gain with #RAG
Register now👉https://t.co/VnF9LAyN5X

[Fangrui Liu @mpsk_liu]

(2/2) Now the chatbot is able to choose WHEN and WHERE to use search, and also the search query is compiled with LLM (actually self-querying retrievers and vector SQL retrievers)

[Fangrui Liu @mpsk_liu]

We just made this showcase on @huggingface space for everyone who want to build a RAG enabled chatbot with @langchain and @MyScaleDB.

[Fangrui Liu @mpsk_liu]

We just made this showcase on @huggingface space for everyone who want to build a RAG enabled chatbot with @langchain and @MyScaleDB.

[MyScale @MyScaleDB]

🚀 Exciting announcement! ChatData just got a major boost! You can now enjoy 💬“chatting” with @arxiv and Wikipedia in addition to the awesome ask/query feature.
Built purely with @streamlit 👑 , @langchain 🦜🔗 and @MyScaleDB and love ❤️ for #AI!
📚Stay tuned for our upcoming blog post to learn all about how we pulled it off!
Try it now: https://t.co/cnA4mEVBns

[MyScale @MyScaleDB]

😎It's your playground for building, testing, and sharing RAG systems.
Major features include:
☑️ Maximum flexibility in designing your retrieval system
☑️ Reproducible evaluations with a YAML configuration
☑️ Detailed traceability of accuracy, execution times, and token usage.

[Fangrui Liu @mpsk_liu]

Why RAG and how is RAG?

Would you like to spend extra cents on each query or thousands of dollars on fine-tuning a model you are not sure if it's working properly?

This article gives you insights on gains from RAG
#RAG #LLM #finetuning

[MyScale @MyScaleDB]

😣The world of RAG systems can be uncertain in terms of performance gains.
📚Get insights from @mpsk_liu’s blog post @thenewstack that introduces a simple end-to-end baseline for evaluating a #RAG System.
😉Interested? Dive into the details now👇

[Fangrui Liu @mpsk_liu]

Open and reproducible, this is how we conduct researches and evaluations! We would like to invite YOU to this research, to build, to share and to enjoy 😎

[MyScale @MyScaleDB]

🤩We’re excited to open-source Retrieval-QA-Benchmark - an end-to-end test workbench for RAG systems
☑️It considers different model settings, retrieval pipelines, knowledge base choices, and search algorithms.
🙌Share with us your evaluation result!
https://t.co/1DZeZk0rgw

[Fangrui Liu @mpsk_liu]

RAG is indeed a good direction to improve truthfulness. We are just curious about how much this improvement would be. We are working on more benchmarks on https://t.co/NRyYrf1ais!

[The New Stack @thenewstack]

Discover the Performance Gain with Retrieval Augmented Generation https://t.co/69BrTHgZ1b @mpsk_liu #LLMs #ChatGPT #HuggingFace #GenerativeAI #RAG

[Fangrui Liu @mpsk_liu]

We updated this PR: https://t.co/aEDVCqrQvW to make every self-querying retrievers to support this feature!

[Fangrui Liu @mpsk_liu]

I also love sef-querying retrievers! And it van be even more powerful than we think! You can even wrap a column with SQL functions!

Take a look at this PR: https://t.co/aEDVCqrQvW

@langchain @hwchase17 #LangChain

[Fangrui Liu @mpsk_liu]

EXACTLY! Thats why we build benchmarks WITH retrieval pipeline.

And check our benchmark here 👇
https://t.co/NRyYrf1ais
to find out how retrieval affects end-to-end RAG performance!

[Arthur Câmara @ArthurCamara]

Llamaindex is nice. But this mini-thread show exactly what the main issue with (most) modern “RAG” frameworks. The ignorance of the whole RETRIEVAL part of RAG.
These are NOT new problems. In fact, they have been discussed for decades in, well, Information Retrieval.

[Fangrui Liu @mpsk_liu]

I also love sef-querying retrievers! And it van be even more powerful than we think! You can even wrap a column with SQL functions!

Take a look at this PR: https://t.co/aEDVCqrQvW

@langchain @hwchase17 #LangChain

[Harrison Chase @hwchase17]

Self-query is maybe my favorite capability in langchain!

An awesome advanced retrieval method

Excited to see support for supabase

[Qin Liu 刘勤 @qliu90]

Join us at the @AIconference to hear from our CEO, Dr. Cheng Tai. Stay tuned for his insightful talk on SQL+Vector!

[Fangrui Liu @mpsk_liu]

Finally a place to donate our prompts. Gets ready for some prompts from @MyScaleDB

#vectordatabase #text2sql

[LangChain @LangChain]

🎡 Introducing LangChain Hub 🦜🔗

A place to publish, discover, and try out prompts

We’re particularly excited about a centralized hub’s promise to enable:
-Encoding of expertise
-Discoverability of prompts for a variety of models
-Inspectability
-Cross-team collaboration

🧵

[Fangrui Liu @mpsk_liu]

The color bar is describing the number of generated tokens. So you can see that more tokens generated the more efficiency you get from @huggingface 's TGI.

[Fangrui Liu @mpsk_liu]

Want to know how prompt tokens affects generation time? Take a look at this 👇
More prompts will eat up more time, but not that bad.

Thanks to nice batching strategy implemented in @huggingface 's text-generation-inference:https://t.co/UfXikE2Cje

#LLM #promptengineering #llama2

[Fangrui Liu @mpsk_liu]

We tested the latency with llama2-13b-chat-hf with 2 @nvidia's RTX3090. Probably this is identical to other models as TGI is distributing the attention heads instead of chunking the model. But we will find out soon.

[Fangrui Liu @mpsk_liu]

Want to know how prompt tokens affects generation time? Take a look at this 👇
More prompts will eat up more time, but not that bad.

Thanks to nice batching strategy implemented in @huggingface 's text-generation-inference:https://t.co/UfXikE2Cje

#LLM #promptengineering #llama2

[Fangrui Liu @mpsk_liu]

Want to know how prompt tokens affects generation time? Take a look at this 👇
More prompts will eat up more time, but not that bad.

Thanks to nice batching strategy implemented in @huggingface 's text-generation-inference:https://t.co/UfXikE2Cje

#LLM #promptengineering #llama2

[Fangrui Liu @mpsk_liu]

🎉🎉

[Jacob Lee @Hacubu]

And a big thank you to 🥇BitVoyagerMan, 🥇@mannybreno, ppramesi, @Tetracyl, @__dqbd, 🥇@puigde14, @mpsk_liu, 🥇@willemmulder, karol-f, and 🥇nameczz for your valuable additions!

That's all - if you're interested in contributing, check out this guide:

https://t.co/5CvgGEPPYE

[Fangrui Liu @mpsk_liu]

We worked so hard to expand how you can work with your data. Self-query retrievers and SQLDatabase are really amazing concepts, that’s why we love @langchain so much! ❤️

[Qin Liu 刘勤 @qliu90]

Use @langchain and @MyScaleDB to reduce hallucination in #LLMs: https://t.co/e6Go6TjR6A