Gen AI Cheat Sheet

Deterministic, greedy decoding. Best for code, math, factual extraction, and JSON schema formatting.

Higher entropy and creative variance. Best for brainstorming, creative writing, and open synthesis.

Truncates long-tail low-probability tokens by dynamically sampling top cumulative 90% probability mass.

Penalizes tokens based on presence in generated text; positive values force model to introduce new topics.

Penalizes tokens based on exact occurrence count; positive values reduce verbatim word repetition.

Split raw corpus using recursive character splitting (500–1000 tokens) with 10–20% chunk overlap.

Convert chunks into dense vectors using embedding models (e.g. text-embedding-3-small).

Index embeddings in vector store using Hierarchical Navigable Small World graphs for fast ANN lookups.

Query vector DB via Cosine Distance or Inner Product to fetch Top-K (K=3–5) context chunks.

Inject retrieved context chunks into System/User prompt template before executing generation.

Begin prompt with explicit expert domain role (e.g. 'You are a Principal Kubernetes Architect...').

Provide 2–3 input/output demonstration pairs prior to presenting the target problem query.

Append 'Think step-by-step before producing your final answer' to force reasoning before generation.

Specify explicit JSON schema structure in system prompt and set `response_format: { type: 'json_object' }`.

Softmax((Q * K^T) / sqrt(d_k)) * V — computes pairwise contextual token weights.

Concat(head_1, ..., head_h) * W_O — projects queries/keys/values into parallel sub-spaces.

W = W_0 + (B * A) * (alpha / r) — trains low-rank adapter matrices while keeping base weights frozen.