level-2-intermediate/individual/12-jargon-level-2/README.md

Level 2 Glossary

LoRA / QLoRA / PEFT: Parameter-efficient ways to adapt a model by training small added components; QLoRA combines this with low-bit quantization. DPO: preference-based tuning that teaches a model which response is preferred. Quantization / quantization-aware training: using lower-precision weights, with training designed to preserve quality. KV cache: stored attention state that speeds continued generation. Speculative decoding: a fast small model proposes tokens that a larger model verifies.

Tool use / function calling / tool-form-filling: a model emits validated structured arguments for software actions. Structured outputs / JSON mode: force responses into a schema; JSON mode alone does not guarantee semantic correctness. Temperature / top_p: sampling controls; lower values are more repeatable. Streaming / batching: return tokens incrementally / process many inputs together. Log probabilities: model confidence signals for generated tokens, requiring calibration.

Embedding drift: retrieval changes as document or embedding distributions change. Rerank: reorder retrieved passages using a stronger relevance model. HyDE: create a hypothetical answer/document to improve retrieval. RAG-Fusion: issue several query variants and merge their results. Agent memory: persisted, scoped state an agent may reuse; never an unbounded transcript. Semantic router: route requests by meaning and confidence.

ReAct / CoT / ToT / agent loop: act-observe reasoning / intermediate reasoning / branching search / the control cycle around tools. MCP / ACP: Model Context Protocol / agent communication protocol. Guardrails: constraints on input, output, tools, and policy. LLM gateway: central layer for model access, routing, spend, and logs. Canary / shadow deploy / feature flag: safe test input / run without serving result / runtime on-off control.

Prompt template / prompt registry: reusable parameterized instruction / governed store of versions. Input cost, output cost, cost per token: billing for tokens sent, generated, and unit pricing. Cache read/write / context caching: reuse stored prompt computation, usually for stable context. Extended thinking / reasoning tokens: provider-controlled extra deliberation tokens; evaluate outcomes, not the story.

Agent trace / span / telemetry / OpenTelemetry: record of a run / timed operation / emitted operational data / common tracing standard. Eval harness / offline eval / online eval / human eval: runner / curated pre-release test / production measurement / reviewer scoring. LLM-as-judge / pairwise ranking / Elo / calibration: model grader / compare two outputs / relative rating / align confidence with reality. MTEB / RAGAS: embedding benchmark / RAG evaluation toolkit. Faithfulness / groundedness: claims supported by context / answer tied to approved sources. Hallucination taxonomy: named classes of unsupported output.

Prompt injection / jailbreak / red team / blue team / purple team: untrusted instruction attack / bypass attempt / attack testers / defenders / joint security practice. Distillation: train a smaller model to imitate a stronger one. Fine-tuning: train model weights for a recurring task. Tool use is permissioned software, not autonomy.