Alibaba ACA GenAI flashcards

31 cards. Read the term and the definition here. The deck at the top of this page is optional.

Qwen

Alibaba's family of large language and multimodal models, released in base, chat, code, and math variants across many parameter sizes for hosted and self-hosted use.

Mixture-of-Experts (MoE)

A model architecture where each token routes to a subset of expert subnetworks, so total parameters are large while compute per token stays comparatively low.

Context window

The maximum number of tokens a model can process in a single request, counting both the input prompt and the generated output together.

Temperature

A sampling parameter that scales output randomness; values near zero yield focused, repeatable answers, while higher values increase variety and creative deviation.

Top-p (nucleus) sampling

A decoding method that samples only from the smallest set of tokens whose cumulative probability reaches p, trimming unlikely tokens while preserving diversity.

Hallucination

Confident, fluent model output that is factually wrong or unsupported by any source, commonly reduced by retrieval grounding and explicit instruction constraints.

Qwen-VL

A multimodal Qwen variant that takes images together with text prompts and returns text, enabling tasks such as image question answering and document understanding.

Tokenization

The process of splitting text into subword tokens that the model consumes; token counts differ from word counts and vary by language and script.

Model Studio

Alibaba Cloud's generative AI application platform offering hosted Qwen and third-party model APIs, prompt tooling, knowledge bases, and application orchestration.

OpenAI-compatible endpoint

A Model Studio access mode that mirrors OpenAI's request and response format, letting existing OpenAI-style client code work by changing base URL and API key.

API key authentication

Model Studio requests authenticate by sending an issued API key in the HTTP Authorization header; the key identifies the account and enforces quotas and billing.

Retrieval-augmented generation (RAG)

A pattern that retrieves relevant documents from a knowledge base at query time and inserts them into the prompt so answers are grounded in supplied content.

Chunking

Splitting source documents into smaller passages before embedding and indexing; chunk size and overlap trade retrieval precision against retained context.

Embedding vector

A numeric vector representing text meaning, produced by an embedding model, so that semantically similar passages sit close together for similarity search.

Cosine similarity

A measure of the angle between two vectors, used in vector search to rank stored chunks by semantic closeness to the embedded user query.

Prompt template

A reusable prompt structure with fixed instructions plus placeholders for variable inputs, keeping application behavior consistent across requests.

Function calling

A mechanism where the model emits a structured request to invoke an external tool or API; the host executes it and returns results for the model to continue.

Streaming response

Incremental delivery of generated tokens as they are produced, lowering perceived latency compared with waiting for the complete response.

Platform for AI (PAI)

Alibaba Cloud's end-to-end machine learning platform covering data preparation, interactive notebooks, distributed training, and online model serving.

PAI-DSW

Data Science Workshop: a managed notebook environment on PAI for interactive exploration and model development, provisioned on CPU or GPU instances.

PAI-DLC

Deep Learning Containers: a PAI service for running distributed training jobs, where users specify the container image, instance type, and instance count.

PAI-EAS

Elastic Algorithm Service: a PAI service that deploys trained models as online inference endpoints with autoscaling and versioned service updates.

PAI-Designer

A visual, drag-and-drop pipeline tool on PAI with prebuilt components for data processing, model training, and evaluation without writing orchestration code.

OSS model artifacts

Object Storage Service buckets commonly store training datasets and model checkpoints that PAI training jobs read and serving services load at deployment.

Inference image

A container image containing the runtime and framework needed to load and serve a trained model on PAI-EAS; it must match the model's framework.

Quantization

Reducing numeric precision of model weights, such as FP16 to INT8, to cut memory use and speed inference at some cost in accuracy.

Lingma

Alibaba Cloud's AI coding assistant, available as IDE extensions, providing inline code completion, chat-based generation, explanation, and test generation.

Agentic coding

An assistant mode where the AI plans multi-step changes, edits multiple files, and runs commands, rather than only suggesting single-line completions.

AI agent loop

The repeated cycle of observing context, planning, invoking a tool, observing the result, and iterating until the task goal is satisfied.

Repository context

Code and files the assistant includes in its prompt, such as open files and related modules, which improves suggestion relevance but raises latency and token cost.

Human-in-the-loop review

A control requiring people to inspect and test AI-generated code before merging, because models can produce plausible but logically incorrect implementations.