Decoding ML Notes
Today is about learning.
Here is a list of learning resources I used and filtered in the past months.
It is one of the most helpful content on Vector DBs, RAG, MLOps and LLMs out there.
This weekโs topics:
Pick the right vector DB for your exact use case
4 video lectures on hands-on LLMs
7 steps you have to achieve 100% MLOps maturity
Advanced RAG
Pick the right vector DB for your exact use case
This is the ๐ผ๐ป๐น๐ ๐ฟ๐ฒ๐๐ผ๐๐ฟ๐ฐ๐ฒ ๐๐ผ๐ ๐ป๐ฒ๐ฒ๐ฑ to ๐ฝ๐ถ๐ฐ๐ธ the ๐ฟ๐ถ๐ด๐ต๐ ๐๐ฒ๐ฐ๐๐ผ๐ฟ ๐๐ for your exact ๐๐๐ฒ ๐ฐ๐ฎ๐๐ฒ.
Since ChatGPT made AI cool, besides the millions of ChatGPT posts you got tired of and blocked, you realized that a new type of tool started to hit the scene: Vector DBs.
As vector DBs play a crucial role in most LLM applications, they popped out everywhere.
On this day, there are 37 vector DB solutions that are constantly changing and adding features.
๐๐ฐ๐ธ, ๐ฉ๐ฐ๐ธ ๐ต๐ฉ๐ฆ ๐ฉ**๐ญ ๐ด๐ฉ๐ฐ๐ถ๐ญ๐ฅ ๐ ๐ฑ๐ช๐ค๐ฌ ๐ฐ๐ฏ๐ฆ?
๐๐๐ง๐ ๐๐จ ๐ฌ๐๐๐ง๐ ๐ฉ๐๐ "๐๐๐๐ฉ๐ค๐ง ๐ฟ๐ฝ ๐พ๐ค๐ข๐ฅ๐๐ง๐๐จ๐ค๐ฃ" ๐ ๐๐๐ ๐จ ๐๐ฃ.
It is an effort managed by Superlinked, where they carefully compared all these 37 vector DBs across 29 features, such as:
- License
- GitHub โญ
- support for text, image or struct models
- RAG, RecSys, LangChain or LllamaIndex APIs
- pricing
- sharding
- document size
- vector dims
...and more!
I won't list all 29 features.
You have to check it out to see them for yourself โ
๐ก๐ผ๐๐ฒ: To keep the table updated or add more features, you can contribute to it yourself.
4 video lectures on hands-on LLMs
Want to build your first ๐๐๐ ๐ฝ๐ฟ๐ผ๐ท๐ฒ๐ฐ๐ but don't know where to start?
Here are ๐ฐ ๐๐ฅ๐๐ ๐น๐ฒ๐ฐ๐๐๐ฟ๐ฒ๐, made by
#1. ๐ ๐ข๐ง๐-๐ญ๐ฎ๐ง๐ข๐ง๐ ๐ฉ๐ข๐ฉ๐๐ฅ๐ข๐ง๐ ๐๐จ๐ซ ๐จ๐ฉ๐๐ง-๐ฌ๐จ๐ฎ๐ซ๐๐ ๐๐๐๐ฌ
You will learn:
- What is model fine-tuning?
- Why is it useful?
- When to use it?
- Why to fine-tune an LLM using QLoRA
- How to architect a fine-tuning pipeline in a real-world project
#2. ๐๐๐ง๐๐ฌ-๐จ๐ง ๐๐ข๐ง๐-๐ญ๐ฎ๐ง๐ข๐ง๐
Let's apply what we learned in lesson 1 to build our first fine-tuning pipeline.
#3. ๐๐ฎ๐ข๐ฅ๐ & ๐๐๐ฉ๐ฅ๐จ๐ฒ ๐ ๐ซ๐๐๐ฅ-๐ญ๐ข๐ฆ๐ ๐ฌ๐ญ๐ซ๐๐๐ฆ๐ข๐ง๐ ๐ฉ๐ข๐ฉ๐๐ฅ๐ข๐ง๐
You will learn:
- How to transform HTML docs into vector embeddings.
- How to process data in real-time
- How to store & retrieve embeddings from a vector DB
- How to deploy it to AWS.
#4. ๐๐ง๐๐๐ซ๐๐ง๐๐ ๐ฉ๐ข๐ฉ๐๐ฅ๐ข๐ง๐
Finally, you will learn how to use LangChain to glue together your fine-tuned LLM and your financial news stored as embeddings in a vector DB to serve predictions behind a RESTful API.
7 steps you have to achieve 100% MLOps maturity
One of the most ๐ฐ๐ผ๐ป๐ณ๐๐๐ถ๐ป๐ด ๐๐ผ๐ฟ๐ฑ๐ in the ๐ ๐ ๐๐ผ๐ฟ๐น๐ฑ is "๐ ๐๐ข๐ฝ๐", a new & interdisciplinary process that isn't fully defined yet.
The good news is that there is a strong movement in ๐ฑ๐ฒ๐ณ๐ถ๐ป๐ถ๐ป๐ด a ๐ฐ๐น๐ฒ๐ฎ๐ฟ ๐๐๐ฟ๐๐ฐ๐๐๐ฟ๐ฒ in ๐๐ฐ๐ผ๐ฟ๐ถ๐ป๐ด the ๐น๐ฒ๐๐ฒ๐น of ๐ ๐๐ข๐ฝ๐ ๐บ๐ฎ๐๐๐ฟ๐ถ๐๐ within your ๐ผ๐ฟ๐ด๐ฎ๐ป๐ถ๐๐ฎ๐๐ถ๐ผ๐ป or ๐ฝ๐ฟ๐ผ๐ท๐ฒ๐ฐ๐.
โณ Here are ๐ณ ๐๐๐ฒ๐ฝ๐ you have to ๐ฐ๐ต๐ฒ๐ฐ๐ธ to ๐ฎ๐ฐ๐ต๐ถ๐ฒ๐๐ฒ ๐ญ๐ฌ๐ฌ% ๐ ๐๐ข๐ฝ๐ ๐บ๐ฎ๐๐๐ฟ๐ถ๐๐ โ
No one other than
๐๐ฒ๐ฟ๐ฒ ๐๐ต๐ฒ๐ ๐ฎ๐ฟ๐ฒ โ
=== ๐๐ถ๐ด๐ต ๐ฉ๐ข๐ท๐ฆ๐ด ===
๐ญ. ๐๐ผ๐ฐ๐๐บ๐ฒ๐ป๐๐ฎ๐๐ถ๐ผ๐ป: project, ML model, and technical documentation
๐ฎ. ๐ง๐ฟ๐ฎ๐ฐ๐ฒ๐ฎ๐ฏ๐ถ๐น๐ถ๐๐ ๐ฎ๐ป๐ฑ ๐ฟ๐ฒ๐ฝ๐ฟ๐ผ๐ฑ๐๐ฐ๐ถ๐ฏ๐ถ๐น๐ถ๐๐: Infrastructure traceability and reproducibility (versioned IaC under CI/CD) and ML code traceability and reproducibility (versioned code, data, and models along with metadata & lineage attached to the data & model)
๐ฏ. ๐๐ผ๐ฑ๐ฒ ๐พ๐๐ฎ๐น๐ถ๐๐: infrastructure code & ML model code quality requirements (tests ran on PRs under the CI pipeline, PR reviews, formatting checks)
๐ฐ. ๐ ๐ผ๐ป๐ถ๐๐ผ๐ฟ๐ถ๐ป๐ด & ๐๐๐ฝ๐ฝ๐ผ๐ฟ๐: infrastructure, application, model performance, business KPIs, data drift and outliers monitoring
=== ๐๐ฆ๐บ๐ฐ๐ฏ๐ฅ ๐ฃ๐ข๐ด๐ช๐ค ๐๐๐๐ฑ๐ด ===
๐ฑ. ๐๐ฎ๐๐ฎ ๐๐ฟ๐ฎ๐ป๐๐ณ๐ผ๐ฟ๐บ๐ฎ๐๐ถ๐ผ๐ป ๐ฝ๐ถ๐ฝ๐ฒ๐น๐ถ๐ป๐ฒ๐ & ๐๐ฒ๐ฎ๐๐๐ฟ๐ฒ ๐๐๐ผ๐ฟ๐ฒ: all the features are shared & versioned from a central feature store
๐ฒ. ๐ ๐ผ๐ฑ๐ฒ๐น ๐๐ ๐ฝ๐น๐ฎ๐ถ๐ป๐ฎ๐ฏ๐ถ๐น๐ถ๐๐: a human can understand the reasoning of the model and not treat it as a black box
๐ณ. ๐/๐ ๐๐ฒ๐๐๐ถ๐ป๐ด & ๐ณ๐ฒ๐ฒ๐ฑ๐ฏ๐ฎ๐ฐ๐ธ ๐น๐ผ๐ผ๐ฝ: inputs & outputs of the model are stored automatically and A/B testing is performed regularly
.
โณ Check out the entire questionnaire on the blog: ๐ MLOps maturity assessment
What level of MLOps maturity is your organization at? For now, you will rarely see 100%.
Advanced RAG
RAG systems are far from perfect โ This free course teaches you how to improve your RAG system.
I recently finished the ๐๐ฑ๐๐ฎ๐ป๐ฐ๐ฒ๐ฑ ๐ฅ๐ฒ๐๐ฟ๐ถ๐ฒ๐๐ฎ๐น ๐ณ๐ผ๐ฟ ๐๐ ๐๐ถ๐๐ต ๐๐ต๐ฟ๐ผ๐บ๐ฎ free course from DeepLearning.AI
If you are into RAG, I find it among the most valuable learning sources.
The course already assumes you know what RAG is.
Its primary focus is to show you all the current issues of RAG and why it is far from perfect.
Afterward, it shows you the latest SoTA techniques to improve your RAG system, such as:
- query expansion
- cross-encoder re-ranking
- embedding adaptors
I am not affiliated with DeepLearning.AI (I wouldn't mind though).
This is a great course you should take if you are into RAG systems.
The good news is that it is free and takes only 1 hour.
Check it out โ