Alibaba ACA GenAI key terms

A neural network trained on vast text corpora to predict and generate language, enabling tasks such as summarization, translation, and question answering through next-token prediction.

A neural network design built on self-attention layers that lets a model weigh the relevance of every input position at once, forming the foundation of modern generative language models.

A computation that assigns weights to different parts of an input sequence so the model can focus on the most relevant information when producing each output.

The basic unit of text a model processes, roughly a word fragment; prompts and outputs are measured and billed in tokens rather than characters.

The preprocessing step that splits raw text into tokens using a fixed vocabulary, converting language into the integer IDs a model can consume.

A dense numeric vector representing text, images, or other content so that semantically similar items sit close together in vector space, enabling search and clustering.

The maximum number of tokens a model can consider in a single request, covering the prompt, retrieved material, conversation history, and generated output combined.

The process of running a trained model to produce outputs for new inputs, as opposed to training, which adjusts the model's weights.

Further training of a pretrained model on a smaller, task-specific dataset so it adapts to a particular domain, style, or instruction format.

Fine-tuning a pretrained model on labeled input-output pairs so it learns to follow instructions and produce responses in a desired format.

A post-training technique that uses human preference rankings to train a reward model, which then guides the language model toward more helpful and safer responses.

A family of fine-tuning methods that update only a small number of added or selected parameters, cutting compute and storage costs while retaining most task gains.

A parameter-efficient technique that freezes the base model and trains small low-rank matrices inserted into its layers, producing compact adapters that can be merged or swapped.

An architecture where each token is routed to a small subset of specialized sub-networks, giving large model capacity while keeping per-token compute comparatively low.

Reducing the numeric precision of model weights and activations, such as from 16-bit to 8-bit or 4-bit, to shrink memory use and speed inference with limited accuracy loss.

A confident but incorrect or fabricated model output, often arising when the model lacks grounding or relevant knowledge for the request.

The process of steering a model's behavior to match human intent and safety expectations, covering helpfulness, honesty, and harm avoidance.

A model that accepts and reasons over more than one data type, such as text plus images, audio, or video, rather than text alone.

Alibaba's series of generative AI models spanning different sizes and modalities, offered in base, instruct, and specialized variants for tasks from chat to coding.

A model whose trained parameters are released for download and local use, allowing self-hosting and customization, though often under a license with usage terms.

The practice of designing and refining instructions, examples, and context so a model reliably produces the desired output format and quality.

A high-priority instruction that sets a model's role, tone, constraints, and safety rules for a conversation, applied before user messages.

Including a handful of worked examples in the prompt so the model can infer the desired pattern, style, or output structure without retraining.

A prompting approach that asks the model to reason through intermediate steps before answering, improving accuracy on multi-step problems.

Alibaba Cloud's managed platform for accessing generative AI models, building applications, and orchestrating workflows without managing underlying infrastructure.

A secret credential that authenticates an application's requests to a model service, typically passed in a request header and rotated periodically for security.

A control that caps how many requests or tokens a client may consume per unit of time, protecting services from overload and enforcing usage quotas.

A pattern that retrieves relevant documents from a knowledge base and supplies them to the model as context, grounding answers in private or current information.

A store optimized for similarity search over embeddings, returning the passages most semantically related to a query vector.

Splitting source documents into smaller passages before embedding and indexing, balancing retrieval precision against the amount of context per chunk.

An AI system that uses a model to plan, choose tools, and take multi-step actions toward a goal, observing results and iterating rather than answering once.

A capability where the model emits a structured request to invoke an external function or API, whose result is fed back so the model can continue reasoning.

The coordination of models, prompts, tools, and data flows in a multi-step application, defining order, branching, and error handling across components.

A configurable control that filters, blocks, or rewrites model inputs and outputs to enforce safety, compliance, and brand policies.

The time between sending a request and receiving a response, a key user-experience metric influenced by model size, output length, and serving infrastructure.

The volume of requests or tokens a serving system can process per unit of time, traded off against latency when sizing deployments.

Publishing a trained model to a serving endpoint so applications can send inference requests to it, with configured resources, scaling, and access controls.

Alibaba Cloud's machine learning platform providing compute, notebooks, training jobs, and serving tools for developing and deploying models at scale.

The curated dataset used to fit or fine-tune a model, whose quality, coverage, and labeling directly shape the resulting model's behavior.

Alibaba's AI coding assistant that provides code completion, generation, explanation, and repository-aware chat inside the developer's IDE and enterprise workflows.