Prompt
The instruction, question, or context you give a model. Good prompts do not need magic words, but they usually benefit from clear intent, relevant context, and an explicit output format.
← About Signal Lens
Glossary
Common AI and ML terms, explained in plain language
This is a selective glossary for terms that appear often across Signal Lens. It is meant to lower the friction of reading the site, not to become an encyclopedia.
Use it when a page assumes you already know the language. Then jump into the linked concept, guide, tool, or model pages when you want the deeper version.
Use this page well
- Read definitions quickly and keep moving if all you need is orientation.
- Follow the related links when the term connects to a bigger system or workflow.
- Return to Start Here if you want a fuller route through the site.
Quick jump
Browse by group
The glossary is grouped by how these terms tend to show up in real conversations, systems, and workflows.
Inputs and context
Terms that shape what goes into a model and how much working room it has.
Tokens
The chunks of text a model actually reads and generates. Tokens are not the same thing as words, so pricing and context limits are usually measured in tokens rather than characters.
Context window
The amount of information a model can consider in one request. A larger context window lets you include more instructions, examples, or source material, but it still needs careful prompt budgeting.
Multimodal
A multimodal model can work with more than one input or output type, such as text, images, audio, or video. In practice that means the same system can often interpret and generate across several media formats.
Embeddings
Numeric representations of meaning that make it easier to compare similarity between pieces of content. They are a core building block in semantic search, retrieval, clustering, and recommendation systems.
Retrieval and grounding
Terms that appear when AI answers need external knowledge instead of memory alone.
RAG
Short for retrieval-augmented generation. It means fetching relevant outside information first and then asking the model to answer using that material, which helps when accuracy depends on current or domain-specific knowledge.
Vector database
A database optimized for storing and searching embeddings. It helps systems retrieve content that is semantically similar, even when the exact wording is different.
Chunking
The practice of splitting source material into smaller pieces before indexing or retrieval. Good chunking keeps each piece focused enough to retrieve well while still preserving useful context.
Reranking
A second-pass ranking step that reorders retrieved items so the most relevant passages surface first. It is often used when initial retrieval finds roughly related content but not the best evidence.
Hallucination
A confident-sounding answer that is wrong, unsupported, or invented. Hallucinations become more likely when the prompt is unclear, the model lacks grounding, or the system has no reliable source of truth.
Orchestration and control
Terms that show up when AI stops being only a chat box and starts acting inside workflows.
Tool use
A model deciding when it should call an outside function, API, or tool instead of answering from text alone. This is what lets models search, fetch data, trigger actions, or operate inside software workflows.
Agent
A system that uses a model plus instructions, tools, and loop logic to pursue a goal across multiple steps. The useful question is usually not whether something is an agent, but how much autonomy and control it has.
Guardrails
The rules, checks, and boundaries wrapped around a model or agent so it stays within acceptable behavior. Guardrails can cover output format, tool permissions, safety rules, review steps, or escalation paths.
MCP
Short for Model Context Protocol. It is a standard way for AI tools to connect to external systems and expose structured capabilities, context, or actions without hard-coding each integration from scratch.
Models and adaptation
Terms that matter when choosing a model, running one, or customizing it for a task.
Inference
The act of running a trained model to produce an output. Training teaches the model, while inference is the day-to-day act of using it.
Fine-tuning
Training an existing model further on new examples so it behaves differently or learns a narrower task. It can improve specialization, but it is usually not the first answer for every quality problem.
LoRA / PEFT
Parameter-efficient ways to adapt a model without retraining every weight. These methods reduce cost and make fine-tuning more practical for teams that want targeted changes instead of full retraining.
Open-weight model
A model whose weights are available for others to download or run under a license. Open-weight does not always mean fully open-source, but it usually gives more control over where and how the model runs.
Proprietary model
A model that is accessed through a provider-controlled product or API without releasing the weights. Proprietary models can be easier to adopt quickly, but they trade away some control and deployment flexibility.