Concept

Open-source AI concepts

Plain definitions of the terms that come up in open-source AI projects, each linked to the projects in our catalogue that implement them.

What is Self-hosting?Self-hosting means running the software on hardware you control instead of renting it, and what it really moves is responsibility: the security patches, the process that has to stay up, the backup that is only real once it has been restored, and the certificate. That is the trade, and the projects here are best read as a spectrum of how much of it one person can reasonably own.What is Semantic search?Semantic search matches a query to documents by meaning rather than by shared words: a model encodes the query and each document into a vector, and the nearest vectors are returned. That is the whole idea, and nearly everything that goes wrong with it comes from the two places the idea hides work, which are the chunking you do before embedding and the index you search afterwards.What is Speculative decoding?Speculative decoding makes autoregressive generation faster by having a small draft model propose several tokens at once and letting the large model verify them in a single batched pass, keeping only the tokens it would have produced itself. The open-source work ranges from training frameworks for draft models to Apple Silicon ports, and the gain depends almost entirely on how often the drafts turn out to be right.What is Speech-to-text (ASR)?Speech-to-text, usually called ASR, turns a waveform into text, and the modern version is a neural sequence model that reads short overlapping windows of audio and emits characters or word pieces directly instead of classifying individual sounds. Everything else in a transcription system, the word timings, the speaker labels and the streaming, is a separate model wrapped around that one.What is Structured output?Structured output (often called JSON mode) is a way of making a language model return data in a fixed, machine-readable shape instead of free prose. The model is constrained by a schema or format so the response can be parsed by code without guessing.What is Text-to-speech (TTS)?Text-to-speech, or TTS, turns text into a waveform, and the modern version does it with a neural acoustic model that predicts speech directly from phonemes rather than stitching together predicted durations and a separately trained vocoder. The reason it now clones a voice from a few seconds of audio is the same change: the model conditions on a reference clip instead of on a voice identifier from a fixed catalogue.What is Tokenizer?A tokenizer is the component that splits raw text or audio into the small units (tokens) a model can process, mapping each unit to an integer ID. Tokenization is the step that turns human-readable input into the sequence a neural network actually consumes.What is Transformer?A Transformer is a neural network architecture built from repeated blocks of self-attention followed by a position-wise feed-forward layer, and that arrangement is what lets a model train over a whole sequence in parallel instead of one token at a time. The open-source ecosystem around it divides into a layer that standardises model definitions, frameworks for vision and post-training, runtimes that execute those models in a browser, and one architecture designed to avoid the attention cost altogether.What is Vector database?A vector database (also called a vector store) stores data as high-dimensional vectors and returns the nearest neighbours to a query vector, usually by approximate search. It is the retrieval layer behind semantic search, recommendation and retrieval-augmented generation.What is VRAM?VRAM is the dedicated memory attached to a GPU, and it is the number that decides whether a model runs on your machine at all, because what a model needs is the sum of its weights, its cache and whatever the framework is holding alongside. The projects here treat that memory as a budget to be scheduled, streamed or avoided, and read together they are the clearest account of what the budget is spent on.