# NewStack.ai

> The AI stack, explained. Plain-English encyclopedia entries mapping every layer of the modern AI stack: models, retrieval, agents, inference, evals, coding, and safety.

## Home

- [NewStack.ai - The AI Stack Explained From A to Z](https://newstack.ai/): Plain-English encyclopedia entries covering LLMs, RAG, agents, MCP, inference, evals, and every other layer of the modern AI stack, from A to Z.

## Core Concepts

- [Core AI Concepts - LLMs, Tokens, and Transformers](https://newstack.ai/concepts/): The vocabulary of modern AI in plain English: how models are built, trained, and run - from tokens and embeddings to fine-tuning and reasoning.
- [What Is an LLM? Large Language Models Explained](https://newstack.ai/concepts/llm/): A plain-English guide to how large language models work: training on text, predicting tokens, and why scale changed what software can do.
- [The Transformer Architecture Behind Modern AI](https://newstack.ai/concepts/transformer/): How attention lets a neural network weigh every word against every other, and why this one architecture powers nearly every modern AI model.
- [Tokens in AI - How Language Models Read Text](https://newstack.ai/concepts/token/): Language models read and write in tokens, not words. What a token is, why it drives pricing and limits, and how tokenization shapes behavior.
- [Context Windows - What Fits in a Model's Memory](https://newstack.ai/concepts/context-window/): Everything a model can consider at once - instructions, documents, conversation - has to fit in the context window. How it works and fills up.
- [Embeddings - How AI Turns Meaning Into Numbers](https://newstack.ai/concepts/embedding/): A vector of numbers that captures what a piece of text means, so similar ideas land near each other. The foundation of semantic search and RAG.
- [Prompt Engineering - Writing Instructions AI Follows](https://newstack.ai/concepts/prompt-engineering/): The craft of structuring what you tell a model - instructions, examples, context, and format - so it reliably produces the output you want.
- [System Prompts - How AI Apps Set Model Behavior](https://newstack.ai/concepts/system-prompt/): The hidden instructions an application sends before any user message, defining the model's role, rules, tone, and tools for the conversation.
- [Temperature in AI - Controlling Model Randomness](https://newstack.ai/concepts/temperature/): A sampling setting that decides whether a model always picks the safest next token or takes creative risks. When to turn it up, down, or off.
- [AI Hallucinations - Why Models Make Things Up](https://newstack.ai/concepts/hallucination/): Language models produce fluent, confident text that is sometimes flat wrong. Why fabrication is built into how they work, and what reduces it.
- [AI Inference - What Happens When a Model Runs](https://newstack.ai/concepts/inference/): Training builds a model; inference is using it. How a request becomes tokens on GPUs, and why speed and cost live in this layer of the stack.
- [Fine-Tuning AI Models on Your Own Data](https://newstack.ai/concepts/fine-tuning/): Continuing a pretrained model's training on your own examples to change its behavior, when that beats prompting or RAG, and when it does not.
- [RLHF - How Human Feedback Shapes AI Behavior](https://newstack.ai/concepts/rlhf/): Reinforcement learning from human feedback trains models toward answers people prefer. How reward models work and why chatbots feel helpful.
- [Quantization - Shrinking AI Models to Run Faster](https://newstack.ai/concepts/quantization/): Storing model weights in fewer bits cuts memory use and speeds up inference at a small quality cost. How it works and why local AI depends on it.
- [Model Distillation - Small AI That Learns From Big](https://newstack.ai/concepts/distillation/): Training a small model to imitate a large one's outputs, keeping most of the capability at a fraction of the cost. Why fast, cheap models exist.
- [Mixture of Experts - Big Models, Fewer Active Parts](https://newstack.ai/concepts/mixture-of-experts/): An architecture that routes each token to a few specialist subnetworks instead of the whole model, buying capacity without the full compute bill.
- [Multimodal AI - Models That See, Hear, and Read](https://newstack.ai/concepts/multimodal/): How models handle images, audio, and video alongside text by mapping every input into the same token space, and what that unlocks in practice.
- [Reasoning Models - AI That Thinks Before Answering](https://newstack.ai/concepts/reasoning-models/): Some models spend extra tokens working through a problem step by step before replying. What that buys, what it costs, and when to reach for one.
- [Open-Weights Models - AI You Can Download and Run](https://newstack.ai/concepts/open-weights/): Anyone can download, run, and fine-tune a model whose weights are published. What open weights does and does not mean, and why it matters.

## Models & Providers

- [AI Model Families and Providers - Claude, GPT, Llama](https://newstack.ai/models/): The major model families - who builds them, open versus closed weights, and how developers access them via APIs, cloud platforms, and local runtimes.
- [Claude - Anthropic's AI Model Family Explained](https://newstack.ai/models/claude/): Anthropic's family of language models, known for long-context work, careful instruction following, and coding. Closed weights, accessed by API and cloud.
- [GPT - OpenAI's Language Model Family Explained](https://newstack.ai/models/gpt/): OpenAI's model family that popularized the modern chatbot era. Closed weights, reached through OpenAI's API, Azure, and the ChatGPT product line.
- [Gemini - Google's Multimodal AI Model Family](https://newstack.ai/models/gemini/): Google's model family, built multimodal from the start and woven through Google Cloud, Workspace, and Android. Closed weights, served via API and Vertex AI.
- [Llama - Meta's Open-Weight AI Model Family](https://newstack.ai/models/llama/): Meta's family of open-weight models that developers can download, run locally, and fine-tune. The default base for much of the open model ecosystem.
- [Mistral - Europe's AI Lab and Its Model Family](https://newstack.ai/models/mistral/): A Paris-based AI lab that ships both open-weight and commercial models, known for compact, efficient releases that run on modest hardware.
- [DeepSeek - Open-Weight Models From China's AI Lab](https://newstack.ai/models/deepseek/): A Chinese lab releasing open-weight models, notable for reasoning ability and training-cost efficiency that reset expectations for what open models can do.
- [Qwen - Alibaba's Open-Weight AI Model Family](https://newstack.ai/models/qwen/): Alibaba's model family, released across many sizes and modalities with open weights, widely used as a base for fine-tunes and multilingual work.
- [AI Model Benchmarks - What They Measure and Miss](https://newstack.ai/models/model-benchmarks/): How standardized tests score language models, why leaderboard rank rarely predicts real-world fit, and how contamination and saturation distort the numbers.
- [How to Choose an AI Model - A Practical Framework](https://newstack.ai/models/choosing-a-model/): A decision framework for picking a model: weigh capability, latency, cost, privacy, and openness against what your application actually needs.

## Retrieval & Data

- [Retrieval and Data - RAG, Vector Databases, Search](https://newstack.ai/retrieval/): How AI systems ground answers in your data: RAG, vector databases, embeddings, chunking, reranking, and the pipelines that feed them all.
- [What Is RAG? Retrieval Augmented Generation Explained](https://newstack.ai/retrieval/rag/): Retrieval augmented generation fetches relevant documents at query time and feeds them to a language model, grounding answers in real sources.
- [Vector Databases - How AI Apps Store and Search Meaning](https://newstack.ai/retrieval/vector-database/): Vector databases store embeddings and answer nearest-neighbor queries, powering semantic search and RAG. Compare pgvector, Pinecone, and Qdrant.
- [Semantic Search - Finding Meaning, Not Just Keywords](https://newstack.ai/retrieval/semantic-search/): Search that matches meaning instead of exact words: queries and documents become embeddings, and results rank by how close their vectors sit.
- [Hybrid Search - Combining Keywords and Vectors](https://newstack.ai/retrieval/hybrid-search/): Why the best retrieval systems run keyword and vector search together, then merge the ranked lists so exact matches and meaning both count.
- [Chunking - Splitting Documents for AI Retrieval](https://newstack.ai/retrieval/chunking/): How documents get split into retrievable pieces, why chunk size and overlap matter, and the strategies that keep meaning intact.
- [Reranking - A Second Pass That Sharpens Search Results](https://newstack.ai/retrieval/reranking/): A cross-encoder scores each candidate against the query and reorders the list, so the best passages reach the model first.
- [Embedding Models - How Text Becomes Searchable Vectors](https://newstack.ai/retrieval/embedding-models/): The models that turn text into vectors for semantic search and RAG: how to compare them, what dimensions mean, and when to switch.
- [pgvector - Vector Search Inside PostgreSQL](https://newstack.ai/retrieval/pgvector/): The Postgres extension that adds a vector type and nearest-neighbor indexes, letting one database hold your rows and your embeddings.
- [Knowledge Graphs - Connecting Facts for AI Retrieval](https://newstack.ai/retrieval/knowledge-graphs/): Entities and relationships stored as a graph give AI systems facts to traverse, complementing vector search when questions span connections.
- [Data Pipelines for AI - From Raw Files to Fresh Indexes](https://newstack.ai/retrieval/data-pipelines/): How raw documents become searchable context: extraction, cleaning, chunking, embedding, and the sync jobs that keep indexes current.

## Agents & Orchestration

- [Agents and Orchestration - How AI Agents Work](https://newstack.ai/agents/): How AI systems act on their own: the agent loop, tool use, MCP, multi-agent patterns, memory, guardrails, and the humans who supervise it all.
- [What Is an AI Agent? Autonomous LLM Systems Explained](https://newstack.ai/agents/ai-agent/): An AI agent pairs a language model with tools and a loop: it observes, decides, acts, and repeats until the job is done. Here is how that loop works.
- [Agentic Workflows - Patterns for Multi-Step AI Tasks](https://newstack.ai/agents/agentic-workflow/): How multi-step AI pipelines chain model calls, branch on results, and retry failures - from fixed sequences to loops the model steers itself.
- [Tool Use and Function Calling in LLMs, Explained](https://newstack.ai/agents/tool-use/): Function calling lets a language model request an action - a search, a database query, an API call - by emitting structured arguments your code runs.
- [MCP Explained - The Model Context Protocol for AI Tools](https://newstack.ai/agents/mcp/): One open protocol connects AI applications to servers that expose tools, resources, and prompts - write an integration once, use it everywhere.
- [Multi-Agent Systems - When One AI Agent Is Not Enough](https://newstack.ai/agents/multi-agent-systems/): Why some AI tasks get split across several agents - orchestrator and worker patterns, handoffs, shared context, and the coordination overhead.
- [Agent Frameworks - Libraries for Building AI Agents](https://newstack.ai/agents/agent-frameworks/): What orchestration libraries and agent SDKs actually handle - the loop, tool wiring, state, retries - and when plain code is the better choice.
- [Agent Memory - How AI Agents Remember Across Sessions](https://newstack.ai/agents/agent-memory/): Context windows forget everything between runs. Memory systems fix that with scratchpads, session summaries, and long-term stores an agent can query.
- [Computer Use - AI Agents That Operate Real Interfaces](https://newstack.ai/agents/computer-use/): Some agents skip the API and drive software the human way - reading screenshots, moving a cursor, typing - so any on-screen app becomes a usable tool.
- [Guardrails for AI Agents - Keeping Outputs in Bounds](https://newstack.ai/agents/guardrails/): Checks that wrap a model's inputs, outputs, and actions - validators, filters, permission gates - so an autonomous system stays inside its limits.
- [Human in the Loop - Where People Fit in AI Automation](https://newstack.ai/agents/human-in-the-loop/): Autonomous does not mean unsupervised. Approval gates, review queues, and escalation paths keep a person on the decisions that matter most.

## Inference & Serving

- [Inference and Serving - Where AI Models Actually Run](https://newstack.ai/inference/): Hosted APIs, cloud GPUs, and local machines: how trained models serve real traffic, with entries on vLLM, Ollama, batching, caching, and streaming.
- [Inference Providers - Hosted APIs for Running AI Models](https://newstack.ai/inference/inference-providers/): Where hosted model APIs fit against cloud GPU platforms and serverless endpoints, and how to pick a provider for latency, cost, and control.
- [Local LLMs - Running Language Models on Your Own Hardware](https://newstack.ai/inference/local-llm/): Run open-weight models on your own machine for privacy, offline use, and zero per-token cost - what it takes and where it falls short.
- [vLLM - The High Throughput LLM Serving Engine Explained](https://newstack.ai/inference/vllm/): How PagedAttention and continuous batching let one GPU serve many users at once, and when vLLM is the right serving engine for your models.
- [Ollama - Run Open Models Locally With One Command](https://newstack.ai/inference/ollama/): A desktop-friendly way to download and run open-weight language models locally, with a built-in API that mimics the hosted providers.
- [GPUs for AI - Why Graphics Cards Run Language Models](https://newstack.ai/inference/gpu-hardware/): Why inference is memory-bound, what VRAM actually limits, and how to think about GPU choices for serving or running models locally.
- [Token Streaming - Why AI Answers Appear Word by Word](https://newstack.ai/inference/streaming/): Language models generate one token at a time, and streaming shows each one as it lands - cutting perceived latency from seconds to instant.
- [Batching - How AI Servers Handle Many Requests at Once](https://newstack.ai/inference/batching/): Grouping requests so a GPU stays busy: how static and continuous batching trade latency for throughput in language model serving.
- [Prompt Caching - Reusing Context to Cut Cost and Latency](https://newstack.ai/inference/prompt-caching/): Store the computed state of a repeated prompt prefix once and reuse it across requests, cutting both time to first token and per-call cost.
- [Structured Output - Getting Valid JSON From AI Models](https://newstack.ai/inference/structured-output/): Constrained decoding and JSON schemas make model responses machine-readable, so downstream code can parse them without regex or retries.

## Evals & Observability

- [Evals and Observability - Testing AI Systems](https://newstack.ai/evals/): How teams know their AI works: offline evals, LLM judges, tracing and observability, red teaming, prompt regression tests, and token cost tracking.
- [What Are AI Evals? Evaluating LLM Applications](https://newstack.ai/evals/ai-evals/): Evals score an AI system's outputs against defined criteria, from golden datasets to online checks, so teams ship changes on evidence instead of vibes.
- [LLM-as-Judge: Using AI Models to Grade AI Output](https://newstack.ai/evals/llm-as-judge/): One model grades another model's answers against a rubric, making evaluation cheap and scalable - if you correct for the judge's known biases.
- [LLM Observability - Monitoring AI Apps in Production](https://newstack.ai/evals/llm-observability/): Logging, tracing, and scoring live LLM traffic so teams can spot quality drops, latency spikes, and cost surges before users report them.
- [Tracing LLM Calls - Traces and Spans for AI Apps](https://newstack.ai/evals/tracing/): A trace records every step of one AI request - prompts, tool calls, retrievals, latencies - so you can replay exactly what the system did and why.
- [Red Teaming AI Systems - Adversarial Testing](https://newstack.ai/evals/red-teaming/): Structured adversarial testing that probes an AI system for jailbreaks, data leaks, and harmful outputs before real attackers find them first.
- [Prompt Testing - Regression Tests for AI Prompts](https://newstack.ai/evals/prompt-testing/): Treat prompts like code: run every change against a fixed test set and compare scores, so a small wording tweak cannot silently break production.
- [Token Cost Tracking - Managing AI Spend per Call](https://newstack.ai/evals/cost-tracking/): Counting input and output tokens per request and pricing them per model turns a surprise AI bill into a metric you can watch, alert on, and cut.

## AI Coding

- [AI Coding - Agents, Vibe Coding, and Modern Workflows](https://newstack.ai/coding/): How software gets built with AI: coding agents, vibe coding, spec-driven development, code review, and the conventions that keep agents on track.
- [AI Coding Agents - How Autonomous Coding Tools Work](https://newstack.ai/coding/ai-coding-agents/): Coding agents read a codebase, edit files, run tests, and iterate until a task is done. What they are, how they work, and where they still fail.
- [What Is Vibe Coding? Prompt-First Software Building](https://newstack.ai/coding/vibe-coding/): Describing what you want and letting an AI write the code, judging results by behavior instead of reading every line. Where it shines and where it bites.
- [Spec-Driven Development for AI Coding Workflows](https://newstack.ai/coding/spec-driven-development/): Writing a precise spec first and letting AI agents implement against it - the discipline that turns vague prompts into reviewable, testable work.
- [AI Code Review - Automated Feedback on Pull Requests](https://newstack.ai/coding/ai-code-review/): How AI reviewers scan diffs for bugs, security issues, and style drift before a human looks - and why they complement rather than replace human review.
- [AI Pair Programming - Coding With an Assistant](https://newstack.ai/coding/pair-programming/): Working side by side with an AI assistant that completes, explains, and refactors code while you stay in the loop on every change.
- [Context Management for AI Coding Agents Explained](https://newstack.ai/coding/context-management/): Why coding agents degrade as context fills up, and the techniques that keep them sharp: compaction, memory files, subagents, and focused sessions.
- [AGENTS.md and Instruction Files for Coding Agents](https://newstack.ai/coding/agents-md/): A plain Markdown file at the repo root that tells AI coding agents how your project works - build commands, conventions, and rules they must follow.
- [AI App Builders - From Prompt to Working Product](https://newstack.ai/coding/app-builders/): Platforms that turn a written description into a deployed application - what they generate, who they serve, and where handwritten code takes over.

## Safety & Governance

- [Safety and Governance - AI Risks, Defenses, Policy](https://newstack.ai/safety/): What can go wrong with AI systems and what teams do about it: safety, alignment, prompt injection, jailbreaking, regulation, and model cards.
- [What Is AI Safety? Risks and Defenses Explained](https://newstack.ai/safety/ai-safety/): The field working to keep AI systems from causing harm, from everyday failures like bad advice to large-scale misuse, and how builders apply it.
- [AI Alignment - Making Models Follow Human Intent](https://newstack.ai/safety/alignment/): How researchers train AI models to pursue what people actually want, why that is hard, and the techniques teams use, from RLHF to constitutional methods.
- [Prompt Injection Attacks on LLM Apps, Explained](https://newstack.ai/safety/prompt-injection/): Attackers hide instructions in content an LLM reads, like web pages or emails, so the model obeys the attacker instead of the developer's instructions.
- [Jailbreaking AI Models - How Attacks Bypass Rules](https://newstack.ai/safety/jailbreaking/): Why users try to talk AI models out of their safety rules, how these attacks differ from prompt injection, and the layered defenses teams deploy.
- [AI Regulation - How Governments Approach AI Risk](https://newstack.ai/safety/ai-regulation/): A builder's map of the regulatory landscape: risk-based frameworks like the EU AI Act, sectoral US rules, and what compliance means for AI products.
- [Responsible AI - Principles Teams Put Into Practice](https://newstack.ai/safety/responsible-ai/): How organizations turn AI ethics into working practice: fairness checks, transparency, human oversight, and governance that survives real deadlines.
- [Model Cards - Documenting How AI Models Behave](https://newstack.ai/safety/model-cards/): The standard document that tells you what an AI model was trained for, where it fails, and how to use it safely - and how to read one before you build.

## A-Z Index

- [A to Z Index - Every AI Stack Term on NewStack.ai](https://newstack.ai/a-z/): Every encyclopedia entry on NewStack.ai in one alphabetical index, from agents and embeddings to vector databases and vibe coding.