Why Chantal ?

Chantal solves the problem of information scattering and lack of identified, authoritative resources, in the context of open-source software projects relying on real-world phenomenons and on the scientific theories describing them. This problem triggers recurring questions from users, coming back in waves on forums and in my emails, and losing many man-hours repeating the same info again and again, as well as server storage and bandwidth, because new posts duplicate content already posted, but never found again.

As image processing is a craft at the crossroad between art, technology and science, it relies on theories and concepts that have precise names, defined by scientific authorities who should be given more credit than random search-engine-optimized blogs. Understanding background theory makes using the image processing tooling a lot less frustrating and a lot more predictable. But users who lack knowledge of the jargon will have no entry point in information systems using keyword-based retrieval methods, which is what blogs, websites and forums use internally for performance.

On the other hand, photography (thus Google index) is polluted by lots of blogs and video channels of people who think they understood color theory because they have been using HSL color wheels in some software for some amount of time. This contributes to spreading false knowledge amongst users and defines a self-feeding loop of time wasting. Things become actively damaging when false knowledge spreads to software developers who use it to support their application design, which has been seen times and times again, especially in the open-source world.

What makes Chantal AI "intelligent" ?

1. The underlying language model knows synonyms and related words: searching for brightness will yield results that are not too far from searching for lightness. This is useful when diving into technical fields without knowing the exact technical slang and to expand your research to similar topics past the anecdotal keywords of your query.
2. The language model knows physical units and computer science: searching for "/usr/lib64/file.so missing" will yield results for _PATH_ _BINARYFILE_ missing, searching for "50mm f/1.8" will yield results for _DISTANCE_ _APERTURE_. This helps generalizing queries and increases the probability of finding relevant pages, even where the exact keywords don't match. Of course, when you know exactly what you are looking for, and don't know where to find it, this smart-ass second-guessing can really get in your way, and this is why you also have an additional grep-like search.
3. The language model uses context to infer topics, so the more keywords you give it, the more likely it is to indentify the right topic of your query.

What makes Chantal AI "eco-friendly" ?

The underlying AI language model is a 512×10000 matrix that weighs about 195 MB and doesn't need more because it is specialized on image-processing topics. It uses an energy-efficient algorithm (Word2Vec, 2013) that turns a search into a matrix dot product. It is trained with around 450k documents containing around 134 millions words, for a total disk space of 560 MB (compressed). Training needs around 6 h on an Intel Xeon laptop. At runtime, searching the whole index of 78k pages needs 3.2 GB of RAM and takes about 0.3 to 0.5 s on a desktop computer.

While this is not the yesteryear state-of-the-art in language processing, the accuracy vs. computational power ratio is very good and lets the search engine run on low-end hardware with reasonable runtimes. The current web service runs on a shared hosting environnement (CPanel/Linux/Apache) typically in 1  s, along with a multi-site WordPress installation and a forum CMS. No GPU is involved.

The current fashion in AI is to run very heavy algorithms (transformers) on GPU farms, and to provide general audience with nice web interfaces allowing them to abuse the service for their own amusement. For information retrieval tasks, these methods are typically 3-6 % more accurate than the one used here, but are at (very) least 500 times more computational-expensive at runtime, while needing much more input training data.

Users should be made aware of the carbon footprint of the cloud-based services they use. Technology is worthless if it solves a problem by creating another one, and increasing electricity consumption is not going to make the world any better.

What makes Chantal better than Google ?

Chantal indexes only manually-curated sources regarding color science and image processing: color theory authorities (CIE, ICC, ACES), researchers (Hunt, Fairchild, Kirk, Poynton), scientific journals (IPOL, Études photographiques), manufacturers (Kodak, Ilford, Nikon), software documentations (Darktable, Ansel, ArgyllCMS) and Github repositiories (for bugs and design decisions). This reduces noise in results and maximizes your probability of finding reliable information, especially since there will be no quiproquo on the meaning of words (technical jargon often reusing common language but with a different meaning – Google is easily confused there).

Chantal also allows you to grep your way through all the indexed document (including PDF), even with regex. Google still doesn't allow that in 2023. Granted, bad regex can put servers down, and we can afford it here because the size of the index is rather small.