Michael Vierling

This is a very interesting article on sparse vectors, which can help overcome limitations of vector search related to domain specific keywords. And of course, the code in the example is clear and clean! Sparse vectors have a few values and many zeros. Dense vectors have limited or no zeroes. But really, what we're talking about semantic search on terms where the context is not present or harder to infer. The author presents this as specialized keyword search such as medical terms or other domain specific terminology. "These scores reflect the significance of each term in the context of the document or query, guiding the model to allocate more weight to terms that are likely to be more meaningful for retrieval purposes." Highly recommend reading. Nice work Paul Iusztin. And a nod to Qdrant and their vector search technology that supports sparse vectors. https://lnkd.in/gcBWfTg8

3

A very interesting article on DSPy and CoT. As complex queries becomes more popular the need to build pipelines is becoming more of a necessity. "When we are building a pipeline, it is not just me trying to convince an LLM to give an output in a certain way, but more of the output should be restricted in such a way that it can work as input for other modules in the bigger pipeline." The shift from prompting to programming to LLM as a module adapting its behavior automatically is going to be very interesting.

9

hype-o-ware (n): A sub-genre of vaporware, distinguished by conformance with current tech hype cycles. e.g., "We'll solve human aging via crypto" or "AI mumble mumble".

64

5 Comments

With the announcement today that Momento's Topics has gone GA, wanted to bring this content back up to the front. If you are looking for rock solid consistency, amazing performance, and the option to bring your own language to connect to their SDK, you need to be checking it out. Here's an article I wrote about Topics and Webhooks with #rust and their SDK https://lnkd.in/ghHuR-cC

13

1 Comment

One of the coolest things we've been building at Truewind is our document parsing and extraction engine.   A lot of accountants deal with scanned or photographed documents like receipts, contracts, invoices, and bills.   For us, the goal is to make sifting through and classifying that information a lot easier.   Matthew Duong has been leading the charge on this project, and recently wrote about the experience on our engineering blog. In particular, he explains why ditching schemas and using PostgreSQL’s JSONB has made such a big difference for our platform.   If you’re interested in the technical side of how we’re refining accuracy and speed in document processing, definitely give it a read.   Here’s the link: https://lnkd.in/gN3Sgext

15

If you are working with Docker... I highly recommend looking at the Slim toolkit. https://lnkd.in/gX_Gb9Kf Why? Slim allows developers to inspect, optimize, and debug their containers using its xray, lint, build, debug, run, images, merge, registry, vulnerability (and other) commands. It simplifies and improves your developer experience building, customizing and using containers. It makes your containers better, smaller and more secure while providing advanced visibility and improved usability working with the original and minified containers. Don't change anything in your container image; minify it by up to 30x to make it secure, too! Optimizing images isn't the only thing it can do, though. It can also help you understand and author better container images. Keep doing what you are doing. No need to change anything. Use the base image you want. Use the package manager you want. Don't worry about hand-optimizing your Dockerfile. You shouldn't have to throw away your tools and your workflow to have small container images. Node.js application images: from ubuntu:14.04 - 432MB => 14MB (minified by 30.85X) from debian:jessie - 406MB => 25.1MB (minified by 16.21X) from node:alpine - 66.7MB => 34.7MB (minified by 1.92X) from node:distroless - 72.7MB => 39.7MB (minified by 1.83X) Go application images: from golang:latest - 700MB => 1.56MB (minified by 448.76X) from ubuntu:14.04 - 531MB => 1.87MB (minified by 284.10X) from golang:alpine - 258MB => 1.56MB (minified by 165.61X) from centos:7 - 615MB => 1.87MB (minified by 329.14X) Rust application images: from rust:1.31 - 2GB => 14MB (minified by 147.16X) Java application images: from ubuntu:14.04 - 743.6 MB => 100.3 MB

4

Tiered storage for Kafka is a classic tarpit idea. It makes all the sense in the world, but it doesn't work in practice. Check out our latest blog post to learn why. https://lnkd.in/eG9BJ95t

107

8 Comments

Gadget's internal CI infrastructure has grown super scaled for a company of our size -- we do 160k hours of CI compute per month, or 220 days of compute per day. We really exhaustively test every change we make across all the framework versions so we don't break your shit! Remarkably though, it doesn't cost us that much. We built a custom autoscaler atop Buildkite that uses GCP Managed Instance Groups on spot instances, using t2d nodes which have really remarkable price/performance and really high spot discounts. Can't speak highly enough of t2ds We decided to use plain old VMs instead of k8s because we spin up a full k3s cluster and integration test our distributed filesystem and our serverless function orchestration stack end to end for some tests. It's caught some serious stuff I resent every moment I have to spend working on CI because I'd rather be building product or yammering on to you all, but I'm glad we made this investment because it is 1/50th the cost of our previous "normal" approach and lets us do nutty stuff like exhaustively test the matrix

54

3 Comments

I ran across this post in the Changelog newsletter: https://lnkd.in/eCdjVekq - I like the name slop for unwanted AI-generated content. I love LLMs. They are part of my daily workflow, and I've been exploring them more seriously while on my sabbatical. I wouldn't misrepresent generated content without critical review as my own, but the web seems to be filled with more generated garbage (or 'slop')—content that looks fine at first but isn't accurate or helpful. LLMs (and other generative AI) should help us to make things of higher quality, not just churn out slop more quickly.

11

1 Comment

“There is no code without a theory of testing.” - Daniel Farina, at the YC Ruby Meetup “If you have 100% branch coverage, it doesn’t mean you’ve covered all the cases. But it does mean that, whenever an obscure fault is understood in production, or even merely observed in development, there’s an incremental path to add it to the base of knowledge in tests: there are no spans of code with no test model.” Daniel has been building infrastructure control planes in Ruby for 13 years - first at Heroku and now at Ubicloud. At the YC meetup, he shared his unusual journey with Ruby and the five features that make Ruby particularly powerful for building infrastructure code. We received solid feedback about Daniel’s talk. So, we edited his years of learnings into a blog post. If you’re interested, please check it out: https://lnkd.in/eHpCjFsK

40

2 Comments