ACP-220 flashcards

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

qwen-turbo vs qwen-max

qwen-turbo is optimized for low latency and low cost, while qwen-max targets the highest reasoning quality and costs more per token with higher latency.

Context window

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

max_tokens parameter

A request parameter that caps the number of tokens the model may generate in its response. It does not limit the input prompt length.

Temperature and top_p

Sampling controls for generation. Temperature scales randomness; top_p restricts sampling to the most probable cumulative mass. Tune one, not both.

Streaming output

Incremental delivery of generated tokens as they are produced, typically over server-sent events. It lowers perceived latency but not total token cost.

Embedding model purpose

An embedding model converts text into dense numeric vectors that capture semantic meaning, enabling similarity search and retrieval for RAG.

Qwen-VL input handling

A multimodal model that accepts images alongside text. Images are encoded into tokens, so higher image resolution increases token usage and cost.

RAG pipeline stages

Parsing, chunking, embedding, vector storage, retrieval, reranking, prompt assembly, and generation. Chunk size and overlap are key tuning knobs.

Chunk size and overlap

Retrieval tuning parameters. Smaller chunks improve precision for specific queries; larger chunks preserve surrounding context but dilute relevance.

Plugin and tool calling

An extension mechanism where the model selects an external API or function, the platform executes it, and the result is returned to the model.

JSON mode output

A generation setting that constrains the model to emit valid JSON, making downstream parsing deterministic for application integration.

System prompt role

The instruction block that sets the assistant's persona, constraints, and output rules before user queries are processed in a conversation.

API key authentication

Requests to Model Studio carry an API key in the Authorization header. Keys should be scoped to least privilege and rotated regularly.

RAG evaluation metrics

Retrieval hit rate measures whether relevant chunks were found; answer faithfulness measures whether the response is grounded in retrieved content.

PAI-EAS

The elastic inference service on Platform for AI. It deploys models as scalable online endpoints with versioning, autoscaling, and traffic splitting.

PAI-DLC

A managed distributed training service on Platform for AI that runs training jobs across clustered GPU resources without manual infrastructure setup.

PAI-DSW

An interactive notebook development environment on Platform for AI used for exploration, experimentation, and prototyping before formal training jobs.

LoRA fine-tuning

A parameter-efficient method that freezes base weights and trains small low-rank adapter matrices, greatly reducing GPU memory and training cost.

Model lifecycle stages

Data labeling, training, evaluation, model registration, deployment, monitoring, and retraining. Each stage gates progression to the next.

Canary deployment

A rollout pattern that sends a small fraction of traffic to a new model version, monitoring metrics before gradually shifting all traffic.

Model drift monitoring

Tracking latency, error rate, and input distribution shifts in production. Significant drift signals the need to retrain or roll back.

Evaluation metric selection

Metrics depend on the task: accuracy and F1 for classification, BLEU and ROUGE for generation, perplexity for language modeling.

AI code completion

Inline, latency-sensitive suggestion of the next code tokens based on surrounding file context, distinct from multi-turn chat-based code generation.

Code prompt specification

Effective coding prompts state the language, framework, input and output contract, error handling expectations, and explicit constraints for the solution.

Verification gates

Unit tests, type checks, and security review applied to AI-generated code before merge, since models can produce plausible but incorrect APIs.

Agentic coding loop

An iterative workflow where the assistant plans, edits files, runs tests, reads failures, and revises, rather than producing a single one-shot output.

Repository context effect

Supplying open files and project conventions to a coding assistant improves relevance and stylistic consistency of generated code.

ASR versus TTS

Automatic speech recognition converts audio into text, while text-to-speech converts written text into synthesized audio for playback.

Streaming ASR

Recognition mode that returns partial transcripts as audio streams in, enabling real-time subtitles; non-streaming returns a single final transcript.

ASR audio preprocessing

Speech recognition commonly expects 16 kHz mono PCM audio, so resampling and channel downmix are standard steps before sending requests.

TTS voice parameters

Text-to-speech requests select a voice plus speaking rate and pitch, and specify output format such as PCM, WAV, or MP3.

Voice pipeline latency

Real-time voice interaction requires streaming end to end: audio capture, ASR, LLM response, TTS synthesis, and playback must each stay within budget.