HNHacker News
TopNewBestAskShowJobs

kacperlukawski

104 karma · joined May 2, 2022

submissionscomments
kacperlukawski··on ChatGPT Plugins
When you develop a plugin, you provide a description that ChatGPT uses to know when to call that particular service. So you don't need to tell people what they need to use - the model will decide independently based on the plugins you enabled.

That being said - we developed a custom plugin for Qdrant docs, so our users will be able to ask questions about how to do certain things with our database. But I do not believe it should be enabled by default for everybody. A non-technical person doesn't need that many details. The same is for the other services - if you prefer using KAYAK over Expedia, you're free to choose.

kacperlukawski··on ChatGPT Plugins
The fact that the model does not have to rely on its internal knowledge anymore but can communicate literally with any external system makes me feel it may significantly reduce the hallucination.
kacperlukawski··on ChatGPT Plugins
That's a game-changer! It seems like factuality issues with ChatGPT might be fixed. We wrote a blog post on how to get started with a custom plugin: https://qdrant.tech/articles/chatgpt-plugin/
kacperlukawski··on Your website's content -> Q&A bot / chatbot
Looks interesting! Have you considered a proper vector database like Qdrant (https://qdrant.tech)? FAISS runs on a single machine, but if you want to scale things up, then a real database makes it a lot easier. And with a free 1GB cluster on Qdrant Cloud (https://cloud.qdrant.io), you can store quite a lot of vectors. Qdrant is also already integrated with Langchain.
kacperlukawski··on From Zendesk to ChatGPT: Leveraging Embeddings
Have you considered using a proper vector database like Qdrant (https://qdrant.tech)? In the demo, it might be just fine to calculate the distance to all the articles in the knowledge base, but in a real case scenario, you'll sooner or later face scaling issues.

Also, libraries such as Langchain are integrated with both LLMs and vector databases, so you can prototype even faster.

kacperlukawski··on On Hybrid Search with Qdrant
The cross-encoder will only rerank the results, that's right. And you're also correct that if the initial search returns 100 garbage results, it won't be able to help. But that's true for any reranking method. Even the rank fusion you use will rerank only the results returned by keyword and vector search. So what is the advantage of it over cross-encoders?
kacperlukawski··on Pgvector: Open-source vector similarity search for Postgres
It is not only about visual similarity, but the semantics of the images as well (for example, all the dog photos should be close to each other, no matter the colours or the scene).

The simplest way to do that is probably to use one of the pretrained neural networks (like resnet), convert the images into embeddings, index them and use for search.

I'm sharing my article describing how to implement it: https://medium.com/p/5515270d27e3

kacperlukawski··on Pgvector: Open-source vector similarity search for Postgres
It would be interesting to include it in the benchmarks with the vector databases: https://qdrant.tech/benchmarks/

Are there any limits except for the max 2,000 dimension that might be indexed?

kacperlukawski··on A Critical Field Guide for Working with Machine Learning Datasets
If you want to perform vector search over your data, then Qdrant (https://qdrant.tech) is worth checking out.
kacperlukawski··on Ask HN: Disk-backed NumPy for data analysis?
For that scale FAISS might be not enough, as it keeps the data in memory and is simply hard to scale. It's worth considering to use a proper vector database in a distributed mode to spread the load, or just allow storing some data on disk: https://qdrant.tech/articles/memory-consumption/
kacperlukawski··on Storing OpenAI embeddings in Postgres with pgvector
What is a memory usage with 40M vectors?
kacperlukawski··on Show HN: I trained an AI model on 120M+ songs from iTunes
Have you considered using a proper vector database, so you can scale it up and run in a cluster? I mean something like Qdrant.
kacperlukawski··on Ask HN: How to build an image search service?
Let me share an article I wrote: https://medium.com/@lukawskikacper/how-to-implement-a-visual... It uses Qdrant and ResNet embeddings to do that. Sources are included if you want to go straight to the implementation.

Qdrant is a Rust-based vector database: https://qdrant.tech/ According to benchmarks that's the fastest vector DB available.

kacperlukawski··on Using LLM to bring out true answers
Since LLMs struggle with providing actual answers, we present how to fix that with a vector database acting as a knowledge base.
kacperlukawski··on Search a million vectors with just 135MB of RAM
Vector databases are supposed to require lots of memory, but this is not true. Vectors might be stored on disks, if they are fast enough, and still keep great performance.
kacperlukawski··on Show HN: I built an AI-powered search engine for a fashion and homeware store
Sure thing! If you decide to use a library that other processes or machines will communicate with, you need to build a whole service around it. The library also doesn't scale well if you need to go beyond a single-machine setup. With a proper database, the distributed deployment is already built-in, so you can scale it up on demand.
kacperlukawski··on Vector Search Engine
Thanks for the insights, really appreciate them! I work at Qdrant, and if there is any way to ease the start with using it, then we're eager to think about it. However, it's way easier to do it with a tool like Stripe, which essentially simplifies one thing. Qdrant should be rather compared to databases like Postgres or MongoDB, but designed to support quick vector search. There is plenty of possibilities of how to use it, including semantic search, image search, anomaly detection, etc. so we don't want to bring a single topic as if it was the main use case. That's why we rather focus on a variety of tutorials and demos, covering different domains.

However, we can surely think of an example based on your data, let's talk! :)

kacperlukawski··on Show HN: I built an AI-powered search engine for a fashion and homeware store
Hey, I wrote an article describing the difference between FAISS, which is a library, and Qdrant - a proper database: https://lukawskikacper.medium.com/i-used-faiss-so-you-dont-h...

TLDR: using a vector database for production systems really pays off, as library is hard to maintain in a long run

kacperlukawski··on Building a semantic search engine in Rust
You're actually comparing a data structure (KD trees) to a library (scann and faiss). HNSW is the data structure that is commonly used by vector search databases. If we're in the Rust world, Qdrant is a great example here.
kacperlukawski··on Will keyword (BM25, TD-IDF) be replaced for search by Neural Search?
Both kinds of systems may coexist - the question that remains open to me is how to combine them both. Vector search, however, provides tools for non-textual search capabilities. Things like querying images, videos, etc, are non-solvable with BM25 and similar methods.
kacperlukawski··on Ask HN: What are the best programming tricks you know?
I completely forgot about that, but that's really one of the ugliest things I have ever seen :D
kacperlukawski··on Ask HN: What are the best programming tricks you know?
That's a massive game-changer, I agree. I really like the approach described by my former colleague: https://softwarephilosopher.com/2022/03/07/personal-worklog/ but I haven't manage to fully implement it
kacperlukawski··on Ask HN: What are the best programming tricks you know?
That's not necessarily a trick I meant, but definitely a good piece of advice. Do you keep them in the codebase after implementation or is it just for you while developing to sort things out?
kacperlukawski··on Similarity Learning lacks a framework. So we built one
Of course, solving a problem is the most important thing at the end of the day. However, there are some data privacy constraints you may sometimes need to fulfil, and sending the data to an external cloud, managed by 3rd party is not an option.

Moreover, fine-tuning might be just one of the applications of neural networks in the organization, and you may already have some pipelines built to train them, so it should be also unified.

And more importantly, Jina's finetuner gives you some pretrained models to choose from, while Quaterion is PyTorch Lightning based, so you can easily integrate it if you already use PyTorch and have the flexibility to fine-tune any custom network as well.

kacperlukawski··on Similarity Learning lacks a framework. So we built one
Unfortunately, it's no longer open source, but requires using their cloud.

"From 0.5.0, Finetuner computing is hosted on Jina Cloud. THe last local version is 0.4.1, one can install it via pip or check out git tags/releases here."

kacperlukawski··on Docker on (rooted) Android
Great! I've just destroyed a screen of my Android mobile and thought it could be great to use it as Raspberry Pi, as fixing that costs more than a new mobile.
kacperlukawski··on Surprising things about HTTP
I used to ask a lot of questions about HTTP during the interviews. It's really surprising how hard backend developers found it to answer those. Even the semantics of the methods, even though REST API were already commonly used.
← PreviousPage 2 of 2