162 articles
Browse guides
Chunking is the least glamorous and most consequential step in a RAG pipeline. Chunks that are too large dilute the embedding and waste context; too small and they lose the meaning that made them relevant. Splitting on structure — headings, sections — beats splitting on character count. In practice: Cutting a contract mid-clause, so retrieval returns half a sentence that means nothing.
A classification model outputs a class rather than a number: spam or not, which of five languages, which disease. Its quality is measured with precision, recall, and related metrics — accuracy alone is misleading whenever one class is rare. In practice: Sorting support tickets into billing, technical, or account.
Computer use lets an agent work with software that has no API by driving the interface directly. It unlocks legacy systems and long-tail tools at the cost of being slower and more brittle than a proper integration. Because it inherits the user’s full access, sandboxing is not optional. In practice: Filling a form in an internal portal that was built in 2009 and has no API.
Conformity assessment is the compliance gate for high-risk systems, done either by the provider itself or through a notified body depending on the use case. It ends in a declaration of conformity, CE marking, and registration in the EU database. Substantial modification can require repeating it. In practice: A documented file showing the risk management, data governance, and testing behind the system. Plain-English summary. The binding text is Regulation (EU) 2024/1689. ...
Content Credentials, built on the C2PA standard, attach a tamper-evident history to a file — captured by this camera, edited here, generated by that model. The approach proves provenance rather than detecting fakes, which is a more tractable problem. Its weakness is that stripping the metadata is trivial, so absence proves nothing. In practice: An image whose credentials show it came from a camera, not a generator.
Context engineering is the discipline that grew up around prompting once context windows got large: with room for hundreds of pages, the question stops being ‘what do I say’ and becomes ‘what does the model actually need’. Irrelevant context measurably degrades output, so curation beats stuffing. In practice: Retrieving three relevant sections instead of pasting the whole 200-page manual.
The context window is the model’s working memory for a single request. Everything must fit: system prompt, chat history, attached files, and the response. Exceed it and the earliest content is dropped or the call fails. Note that a large window is not the same as good recall inside it — models still lose track of the middle. In practice: A 200K window holds roughly a 500-page book — but the answer must fit in there too. ...
A copilot works alongside you inside the application — suggesting the next line, drafting the paragraph, proposing the formula. The distinction from an agent is who holds the wheel: a copilot proposes, you accept. The name is now generic despite also being a product brand. In practice: Autocomplete that finishes the function you were about to write.
Poisoning attacks the model before it exists. A small amount of crafted data can install a backdoor that behaves normally except on an attacker’s chosen input. It is a supply-chain problem, which is why data provenance is a security control and not just documentation. In practice: Injected examples that make a filter approve anything containing a particular phrase.
Provenance answers the questions that arrive with an audit or a lawsuit: whose data is this, on what basis was it collected, and are we allowed to train on it. Retrofitting it is painful and sometimes impossible. It also underpins reproducibility and poisoning defence. In practice: Being able to show that no scraped personal data entered the training set.