The vocabulary of the directory, defined plainly. Every entry on Hyland links back to these concepts — and every definition is written to be quoted.
Every entry in the directory carries a 1–100 score. For GitHub repositories the score is computed — not editorialized — from public signals, refreshed weekly:
For platforms and agent frameworks — where no single public metric tells the story — the score blends community adoption, feature completeness, documentation quality, active development, and ecosystem size, reviewed weekly by our automated pipeline.
A software system that uses a large language model to pursue a goal autonomously: it plans steps, calls tools (APIs, code execution, web search), observes results, and iterates — instead of answering a single prompt. Frameworks like LangGraph, AutoGen, and CrewAI provide the scaffolding for building agents.
An architecture where several AI agents with distinct roles (planner, coder, reviewer) collaborate on a task, passing messages and delegating subtasks. Useful when a problem decomposes into specialties, at the cost of orchestration complexity.
A 1–100 quality ranking computed by Hyland.ai for every directory entry. For GitHub repositories it is fully deterministic — computed from public signals: stars (30%), forks (15%), commit velocity (20%), documentation (15%), community (10%), and freshness (10%). For platforms and agent frameworks it blends community traction, feature completeness, documentation quality, and editorial weighting.
A model whose trained parameters are published for anyone to download, run, and often fine-tune — e.g. Llama, DeepSeek, Qwen. Distinct from "open source" in the strict sense: training data and code may remain closed, and usage is governed by the model license.
Software you can run on your own infrastructure rather than the vendor's cloud. For AI, this means your prompts and data never leave your machines — a hard requirement in regulated industries and the main reason open-weights models matter.
The maximum amount of text (measured in tokens) a model can consider at once — your prompt plus its working memory. Larger windows (128K–1M tokens) enable whole-codebase analysis and long-document reasoning; smaller ones are cheaper and faster.
A pattern where the model retrieves relevant documents from a knowledge base (usually via a vector database) before answering, grounding responses in your private data instead of relying on training memory. The standard approach for enterprise chatbots.
A model architecture that activates only a fraction of its total parameters per token — e.g. 37B active out of 671B total. Delivers large-model quality at small-model inference cost; used by DeepSeek, Qwen3, Llama 4, and Kimi K2.
Further training a pre-trained model on your own data to specialize its behavior, tone, or domain knowledge. Parameter-efficient methods (LoRA, QLoRA) make this possible on a single GPU for open models.
A model's ability to emit structured requests to external tools — search, calculators, APIs, databases — and incorporate the results. The primitive that makes agents possible.
A neural network trained on massive text corpora to predict and generate language. The engine inside chatbots, agents, and most AI tools — GPT-4o, Claude, Llama, and DeepSeek are all LLMs.
The practice of structuring inputs to get reliably better outputs from a model — role assignment, constraints, examples, and step-by-step instructions. The cheapest performance upgrade available.
When a model generates plausible-sounding but false information — invented facts, citations, or code that doesn't exist. The central reliability problem; mitigated by grounding, retrieval, and verification.
Running a trained model to produce outputs — as opposed to training it. Inference cost and latency determine what a model costs to serve, which is why smaller efficient models matter.
The time between sending a request and receiving the first token of the response. Critical for chat, voice, and coding assistants — a model that's smart but slow loses to one that's good and instant.
The unit models read and write — roughly ¾ of an English word. Pricing, context windows, and rate limits are all denominated in tokens.
A numeric vector representation of text that captures meaning, so similar texts land close together. The foundation of semantic search and RAG pipelines.
A database optimized for storing embeddings and finding the nearest neighbors of a query vector. The retrieval layer in most RAG systems — Pinecone, Weaviate, pgvector, Qdrant.
Constraints layered around a model to keep outputs safe, on-topic, or correctly formatted — input filters, output validators, refusal policies. What separates a demo agent from a production one.
A design pattern where an agent pauses for human approval at defined checkpoints — before spending money, sending messages, or executing irreversible actions. The standard for trustworthy automation.
A task executed by a model that plans, uses tools, checks its own results, and iterates — rather than answering in a single pass. The shift from "chatbot" to "coworker".
The standing instruction set that defines a model's role, scope, and behavior before any user message. The single highest-leverage control point in any AI application.
Training a smaller model to imitate a larger one's outputs, keeping most of the capability at a fraction of the cost. How many efficient "small" frontier models are made.
Compressing a model's weights from 16-bit to 8/4-bit precision so it runs on cheaper hardware with minor quality loss. What makes 70B models runnable on a single GPU.
A model that handles more than one data type — text plus images, audio, or video — in a single system. Enables tasks like "describe this screenshot" or "transcribe and summarize this call".