Knowledge base
The AI and cloud security library
Practical, vendor-neutral explainers on AI, API, web, container and cloud security. All 207 are free to read, with no login. New here? Start with path 01, then work across.
Filter by topic (137)
-
Observability & tracing (OpenTelemetry GenAI semantics)
How OpenTelemetry GenAI spans capture prompts, completions, tokens, cost and tool calls in a trace, and why plain logs cannot explain a bad answer.
-
Open vs closed, hosted vs self-hosted models
Compares closed and open-weight models, hosted and self-hosted, so you can pick by quality, data control and cost.
-
Orchestration Frameworks
Shows what an agent orchestration framework runs for you, where it differs from a fixed workflow, and when a plain loop is enough.
-
PII Redaction / Prompt DLP
Explains how to detect and tokenize personal data and secrets in prompts and outputs before they reach an external LLM.
-
Planning & Reasoning Patterns
Compares chain-of-thought, ReAct and tree-of-thought prompting, with the cost of each and when a model needs none.
-
Prompt / version registries
Describes how a prompt registry stores versions, labels and eval scores so prompts can be fetched, tested and rolled back.
-
Prompt engineering (zero/few-shot)
Explains zero-shot and few-shot prompting, why examples steer output without retraining, and when fine-tuning becomes worth it.
-
Prompt templates & prompt management
Covers parameterized prompt templates, safe interpolation, versioning and eval-driven testing, and why inline f-strings fall short.
-
Provider key & secret management
Covers storing, scoping, rotating and monitoring LLM provider API keys, with a gateway and spend limits to contain a leak.
-
Provider-Level Prompt Caching
Explains how provider prompt caching reuses the KV cache for a stable prefix, what invalidates it and when it pays off.
-
Quantization, distillation & model-size tradeoffs
Compares quantization and distillation for shrinking models, what each costs in quality, and how to match model size to the task.
-
RAG (Retrieval-Augmented Generation)
RAG retrieves matching document chunks and places them in the prompt; this article walks the pipeline and where grounding fails.
-
Re-ranking
A cross-encoder re-ranker re-scores the retriever's top candidates so the best passages reach the context window; here is when it pays off.
-
Real-time data & live context
Live context fetches current data at request time and injects it into the prompt; this covers the flow, cache TTLs and stale-data failures.
-
Response caching (exact & semantic)
Exact caching keys on the literal prompt and semantic caching on meaning; this compares both, plus prompt caching, and the false-hit risk.
-
RLHF & DPO (preference tuning)
RLHF trains a reward model then runs PPO, while DPO optimizes preferences directly; this compares both and the risk of narrow preference data.
-
Sampling, temperature & decoding
Temperature, top_p and top_k control how a model picks each next token; this explains the three settings and which values suit which tasks.
-
Semantic Caching
A semantic cache returns a stored answer when a new query embeds close to a past one; here are the questions teams ask about thresholds and fit.
-
Small / On-Device Language Models (SLMs)
Small language models of roughly 1 to 14B parameters run locally when quantized; this covers fit, trade-offs and the data-privacy benefit.
-
Streaming responses (token streaming / SSE)
Token streaming sends model output over a long-lived SSE connection as it is generated; this covers transports, client duties and when to avoid it.
-
Structured Error Handling for Model Calls
An LLM call fails in distinct ways, from 429 rate limits to a 200 response with bad output; this sets out how to route each failure class.
-
Structured output / JSON mode / function schemas
Learn how schema-constrained decoding and function calling force valid JSON, and why you still validate the values.
-
Synthetic data generation
Understand how synthetic training data is generated and filtered, and why unfiltered output can degrade the model you train.
-
System vs user vs developer roles
See how system, developer and user roles rank instructions in a chat API, and why the boundary is a soft one you must back with code.