The DriftRail knowledge base
Understand the system.
Know what to watch.
Field notes on observing, evaluating, and protecting AI in production. Clear answers, practical checks, and the limits that matter.
The reading room / 04
Start with a question.
What is LLM Latency?
LLM latency includes time to first output, generation time, retrieval, tools, and retries. Measure user journeys and latency distributions.
What is LLM Observability?
Understand LLM observability: traces, costs, latency, output evaluation, and the evidence needed to investigate production failures.
What is LoRA?
Learn what LoRA (Low-Rank Adaptation) is and how it enables efficient LLM fine-tuning. Understand why monitoring fine-tuned models matters.
What is MCP?
Learn what MCP (Model Context Protocol) is and how it standardizes AI tool integration. Understand the implications for AI observability and monitoring.
What is Mixture of Experts?
Learn what Mixture of Experts architecture means for LLMs. Understand how MoE enables larger models with efficient inference.
What is Model Collapse?
Learn what model collapse means. Understand how training on AI-generated data degrades model quality and why monitoring matters.
What is Model Drift?
Model drift is a change in an AI system’s observed behavior or performance. Compare stable task groups before attributing a change to the model.
What is Multimodal AI?
Learn what multimodal AI means. Understand how models process text, images, audio, and video, and the unique monitoring challenges.
What is PII Detection in AI?
PII detection identifies potential personal information in prompts, responses, and logs. Learn where to check it and why detection is imperfect.
What is Prompt Engineering?
Learn what prompt engineering means. Understand techniques for crafting effective prompts and why monitoring prompt performance matters.
What is Prompt Injection?
Prompt injection is an attack where malicious inputs manipulate LLM behavior to bypass safety controls, leak data, or perform unauthorized actions. Learn how to detect and prevent it.
What is Prompt Versioning?
Prompt versioning is the practice of tracking, managing, and deploying different versions of prompts used with LLMs. Learn why it matters for production AI systems.
What is LLM Quantization?
Learn what quantization means for LLMs. Understand how reducing model precision affects performance, latency, and why monitoring quantized models matters.
What is a Reasoning Model?
Learn what reasoning models are and how they differ from standard LLMs. Understand o1, o3, and extended thinking capabilities.
What is RAG?
RAG retrieves external information for a model to use while answering. Learn the retrieval pipeline, source checks, and common failure modes.
What is RLHF?
Learn how RLHF works to align AI with human preferences. Understand the training process that makes ChatGPT and other LLMs helpful and safe.
What is Semantic Caching?
Learn what semantic caching means for LLMs. Understand how caching similar queries reduces costs and latency while maintaining quality.
What is LLM Streaming?
Learn how streaming delivers LLM responses token-by-token for better UX. Understand SSE, WebSockets, and streaming implementation.
What is Structured Output?
Learn what structured output is and how LLMs generate valid JSON, XML, and other formats. Understand schema validation and monitoring.
What is Synthetic Data?
Learn what synthetic data means for AI training. Understand benefits, risks, and quality considerations for AI-generated training data.
What is a System Prompt?
Learn how system prompts control LLM behavior. Best practices for writing effective system instructions for AI applications.
What is Temperature?
Learn how temperature controls LLM randomness and creativity. Understand when to use high vs low temperature settings.
What is Tokenization in LLMs?
Learn how tokenization works in large language models, why it matters for costs and context limits, and how different tokenizers handle text.
What is Tool Use in AI?
Learn what tool use and function calling means in LLMs. Understand how GPT-5, Claude 4, and Gemini 3 interact with external APIs and why monitoring matters.