Simon Marache-Francisco

Over the weekend, I implemented a Retrieval-Augmented Generation (RAG) pipeline as an experiment. RAG is an AI-driven document search technique. While it is powerful and interesting, the fundamentals turned out to be a lot more basic than I expected. The source code of the pipeline is available in the microProjects repo. What does RAG or AI Search look like? From the point of view of the user, RAG is a sophisticated document search engine.

Recently, I’ve been working on a few small but fun projects. As usual it’s mostly about software, but there is a bit of (light) embedded too this time! These involve using MicroPython for a hardware project, creating a user-friendly interface for DALL-E, and setting up a self-hosted music streaming solution. USB Volume Knob with CircuitPython I created a volume/media key USB controller for my desktop using MicroPython and a ProMicro microcontroller.

Post in French Introduction I was talking to my father about the Lewis and Clark expedition of 1804. I found the original document (in the public domain) on the gutenberg project’s webpage. I translated a couple of entries for him and he looked interested. Unfortunately I couldn’t find any French translation of these work for him to read. Given the simple nature of the medium (cleanly separated entries with a title), we wondered if we could use ChatGPT to perform the translation.

[Français] Presentation Ceci est une traduction française du livre “The Journals of Lewis and Clark, 1804-1806” qui contient les récits de voyages de William Clark et Meriwether Lewis lors de leur expédition à travers l’Amérique. Je n’ai pas pu trouver de traduction de ce texte en français et j’ai donc décidé d’utiliser GPT-4 Turbo d’OpenAI pour réaliser une traduction automatique. Bien que le résultat soit de très bonne qualité, je ne peux faire aucune promesse quant à la fidélité de la traduction.

Over the year I have been using several platform to host my projects, from S3/cloudfront/ACM, docker-compose/VPS to k8s, back to docker-compose/VPS again. But I recently found out that github pages was supporting artifacts created by github actions and custom subdomains in the free tier. I then migrated most of my static projects: Scoring assistant for Archery: https://score.marache.net Simple hex/dec/binary converter: https://convert.marache.net This website: https://marache.net Recipe generator for Bread-making: https://pain.marache.net That was also a nice opportunity to clean up the remnant of the travis build pipelines, and replace them with github action pipelines.

I am a backend/DevOps, so I haven’t a lot of opportunities to work on any kind of GUI. But at the same time, I am always fascinated by image generation (See my voxel renderer, software rasterizer, and many fractal generators), and having a simple engineering UI makes life much easier for the developer. Since my last project was (another) fractal generator, written in rust/OpenGL this time, I decided to look for a simple UI component that I could use to provide the interface, and stumbled on egui (the UI) + glow (the OpenGL bindings).

I read about the gardner hype cycle a long time ago, but it is only in the last few years that I feel that I fully “got” it. And now, I see it everywhere. There’s always a new technology promising to revolutionize everything, change the world, only to fall short of the (often exaggerated) expectations. As I’ve grown older, I’ve come to realize that the most visible failures often thrive within a VC/media echo chamber, detached from the realities on the ground.

I had a bit of fun over the last few months working on server apps. I focused on building simple, well-contained services that I would be able to use to learn new paradigm and languages: bibin: a swiss knife of a pastebin I started with something simple but useful to learn the rocket framework, SQLite and async programming in rust. I made a fork of bin, a very minimalist pastebin-like service, that I upgraded by adding all the features that I could think of:

I wanted to have a quick look at the PNG implementation, to see if I could have a high level understanding of the W3C spec over the weekend. After a quick implementation of the chunk parsing logic in rust, I started to dig into the specification for the IDAT chunk, … and things got a bit ouf of hand… I am now reading the RFC1950 (zlib data format) and RFC1951 (DEFLATE algorithm), as they are are used to encode the image data.

I made a simple implementation of the md5 algorithm based on the original RFC 1321 from 1992! It was quite interesting to see the original code for the reference implementation, as it is using the old-style K&R syntax. For instance, the types for function arguments are defined separately: /* Decodes input (unsigned char) into output (UINT4). Assumes len is a multiple of 4. */ static void Decode (output, input, len) UINT4 *output; unsigned char *input; unsigned int len; { I tried to stay as close as possible to the RFC, without considerations for speed.