ACP-220 cheat sheet

Model families on Alibaba Cloud Model Studio include Qwen text generation (qwen-max, qwen-plus, qwen-turbo), Qwen-VL multimodal, Qwen-Audio, and Qwen-Coder code models.

qwen-turbo prioritizes latency and cost; qwen-plus balances quality and price; qwen-max targets highest reasoning quality at higher cost and latency.

Context window is measured in tokens; a common rule of thumb is 1 Chinese character ≈ 1 token and 1 English word ≈ 1.3 tokens.

max_tokens caps generated output length, not input length; input plus output must fit inside the model's context window.

temperature controls randomness (higher = more diverse), top_p performs nucleus sampling; adjust one of the two, not both simultaneously.

Streaming (incremental output via SSE) reduces perceived latency but does not change total token billing.

Qwen-VL accepts image plus text input; images are converted to tokens, so image resolution affects token consumption and cost.

Embedding models (text-embedding) return dense vectors used for semantic search and RAG retrieval; they do not generate text.

Model Studio applications are built from prompt templates, knowledge bases (RAG), plugins, and workflow orchestration.

RAG pipeline stages: document parsing → chunking → embedding → vector storage → retrieval → rerank → prompt assembly → LLM generation.

Chunk size and overlap are the primary retrieval tuning knobs; smaller chunks improve precision, larger chunks preserve context.

Plugins extend an application with external APIs or tools; the model selects a tool via function/tool calling and the platform executes it.

API access uses an API key in the Authorization header with DashScope-compatible endpoints; keys should be scoped and rotated.

Structured output can be enforced with JSON mode or function calling schemas to make downstream parsing deterministic.

Prompt templates support variables injected at runtime; system prompts set role and constraints, user prompts carry the query.

Evaluation of an application should track retrieval hit rate, answer faithfulness, and latency alongside token cost.

PAI provides modules for data preparation, training (PAI-DLC), interactive development (DSW), and online inference (PAI-EAS).

PAI-DLC runs distributed training jobs on managed clusters; PAI-DSW is a notebook IDE for exploration and experimentation.

PAI-EAS deploys models as elastic inference services with autoscaling, versioning, and traffic splitting for canary rollout.

Model lifecycle: data labeling → training → evaluation → registration in a model repository → deployment → monitoring → retraining.

Fine-tuning (SFT) adapts a base model to a task; LoRA/QLoRA reduce trainable parameters and GPU memory versus full fine-tuning.

Evaluation metrics depend on task: accuracy/F1 for classification, BLEU/ROUGE for generation, perplexity for language modeling.

Model monitoring tracks latency, QPS, error rate, and data drift; drift signals trigger retraining or rollback.

Training cost scales with GPU-hours; batch size and learning rate are tuned jointly, and mixed precision lowers memory footprint.

AI coding assistants generate code from natural-language intent, complete partial code, and explain or refactor existing code.

Effective prompts for code specify language, framework, input/output contract, error handling, and constraints explicitly.

Generated code must be reviewed and tested; models can produce plausible but incorrect APIs or deprecated method calls.

Unit tests and type checks act as verification gates for AI-generated changes before merge.

Context supplied to the assistant (open files, repository conventions) strongly affects output relevance and style consistency.

Code completion differs from chat-based generation: completion is inline and latency-sensitive, chat supports multi-turn reasoning.

Security review is required because generated code may introduce injection flaws, hardcoded secrets, or unsafe dependencies.

Agentic coding workflows iterate: plan → edit files → run tests → read failures → revise, rather than one-shot generation.

Speech-to-text (ASR) converts audio to text; text-to-speech (TTS) converts text to audio, both exposed as cloud APIs.

ASR accuracy depends on sample rate, channel count, background noise, and domain vocabulary; custom hotwords improve recognition of names.

TTS supports voice selection, speaking rate, and pitch; output audio format (e.g., PCM, WAV, MP3) must match the consumer.

Streaming ASR returns partial results as audio arrives, enabling real-time subtitles; non-streaming returns one final transcript.

Multimodal models accept image, audio, and text; modality tokens count toward context and billing like text tokens.

OCR extracts text from images; for scanned documents, layout-aware parsing improves table and multi-column accuracy.

Audio is typically 16 kHz mono PCM for ASR; resampling and channel downmix are common preprocessing steps.

Latency budgets for voice interaction require streaming pipelines end-to-end: capture → ASR → LLM → TTS → playback.