brainbaby

AI glossary

Look up the terms you see in AI products and research. Each definition has a short explanation and a related guide or reference.

Agentic AI

AI that can plan a task and use tools across several steps.

An assistant could look up calendar availability and draft a meeting invitation. The application decides which tools are available and whether a person must approve an action.

AI App Builder

A no-code tool that generates a working mobile or web app from a natural-language spec.

A mobile app brief should describe screens, data and interactions. Check the preview and test sign-in and on-device behavior before submitting a build.

AI Coding Assistant

An LLM specialized for writing, reviewing, and refactoring source code.

Explore Baby 5.0 for coding and project work in Brainbaby.

AI Document Creator

A tool that drafts long-form business documents from a brief.

A brief can produce a proposal or report draft. Verify names, amounts and commitments before you share the document.

Chain of Thought (CoT)

A technique that uses intermediate steps in solving a problem.

A step-by-step explanation can help a reader check an answer, but it may still contain mistakes. Check important conclusions against the underlying evidence.

Context Window

The amount of text and other input a model can consider in one request, measured in tokens.

Instructions, conversation history and retrieved documents can all take up space. Product limits may be lower than a model’s advertised window. For a long PDF, start with the sections that matter and ask for page references.

Diffusion Model

A generative model that turns noise into images or video through iterative denoising.

For an image, the model can use a prompt to guide denoising toward a subject or style. Different image and video models use different architectures.

E-Signature

An electronic indication that a signer intends to sign a document.

Signature requirements vary by document and jurisdiction. Check the rules and use a signing workflow that records the required evidence.

Embedding

A numeric vector that represents the semantic meaning of a piece of text.

Embeddings can support semantic search, retrieval and memory by making related content easier to compare. Results still need relevance and accuracy checks.

Face Swap

Replacing the face in an image or video with another face while preserving expression.

Use images you are authorized to edit and obtain consent from the people depicted. Review the output before sharing it.

FAQ Schema

A structured JSON-LD block that pairs questions with their canonical answers.

The questions and answers should reflect content visible on the page. Adding FAQ markup does not guarantee a rich result or an AI citation.

Fine-Tuning

Continued training on a smaller dataset to specialize a base model.

Fine-tuning changes model weights. Giving a model instructions or retrieving a document for one request does not change those weights.

GEO (Generative Engine Optimization)

Work to make public content useful and understandable to AI-assisted search and answer systems.

Clear, accurate content and accessible pages can help people and systems understand a source. No file, schema or publication guarantees a citation.

GPQA (Graduate-Level QA)

A benchmark of difficult graduate-level science questions.

GPQA measures whether an LLM can reason about biology, chemistry, and physics beyond memorized facts.

Hallucination

When an LLM confidently states something factually wrong.

A generated answer can invent a citation or state an unsupported date. Open the source and check whether it supports the specific claim.

HumanEval

A benchmark that measures whether an LLM can write correct code for 164 Python problems.

HumanEval is the classic code-generation benchmark.

Image-to-Video

Animating a still image into a short video clip.

A source image sets the starting composition. Specify the motion you want and review the chosen model’s duration, resolution and cost.

Inference

Running a trained model to generate an answer, as opposed to training it.

A product can meter inference by tokens or use other units depending on the task, such as image jobs or video duration.

Knowledge Base

A private collection of documents an AI can search and cite in its answers.

A knowledge base can hold policies or project notes. Check which documents a response cites and whether you have permission to share them.

Large Language Model (LLM)

A model trained on large amounts of data to process and generate language.

Large language models power chat assistants, code generation, and document creation.

Life Graph Memory

Brainbaby's persistent memory model that stores people, projects, and events as a graph.

See the product page for current memory behavior and controls.

llms.txt

An optional text file that can summarize a website and link to useful resources.

Brainbaby publishes /llms.txt and /llms-full.txt with links to public information. Different services may use or ignore these files; they do not guarantee crawling or citations.

Mixture of Experts (MoE)

An LLM architecture that routes each token to a small subset of expert sub-networks.

MoE lets a model scale total parameters far beyond what a dense model of the same inference cost could handle.

MMLU

A 57-subject multiple-choice benchmark for general knowledge.

Compare scores only when the model evaluation uses the same tasks and scoring conditions.

MMLU-Pro

A multiple-choice benchmark designed to test knowledge and reasoning across multiple domains.

MMLU-Pro measures general reasoning across STEM, humanities, and professional exams.

Model Context Protocol (MCP)

An open protocol for connecting AI applications to external tools and data.

A server can expose resources or functions to a compatible application. That application still manages permissions and user approval for tool actions.

Model Distillation

Training a smaller student model to mimic the outputs of a larger teacher model.

A student model learns from examples or outputs produced by a teacher model. Speed and quality depend on the models, data and evaluation used.

Persistent Memory

Information an assistant can reuse across conversations, subject to its settings and storage rules.

A saved writing preference can help in a later chat. Review what the product stores, how to change or delete it, and when it may be applied.

Presentation Maker

An AI tool that builds a full slide deck from a single prompt.

A useful deck starts with the audience, key message and source notes. Review the exported slides for factual errors, layout and missing evidence.

Prompt Engineering

The craft of writing inputs that elicit high-quality LLM outputs.

For an image edit, specify the change and what to preserve. “Replace the background with pale gray and keep the label readable” is easier to check than “make this better.”

Quantization

Compressing model weights to lower precision to shrink memory and speed up inference.

Reducing numerical precision can lower a model's memory needs, sometimes with a quality tradeoff. The result depends on the model and hardware.

Reinforcement Learning from Human Feedback (RLHF)

Training an LLM against human preference rankings to shape its responses.

Feedback can help shape the model’s behavior. Results depend on the feedback, training method and evaluation.

Retrieval-Augmented Generation (RAG)

Finding relevant information at request time and giving it to a model as context for its answer.

A support assistant might retrieve the current returns policy before answering. Check that its answer reflects the cited policy. A saved preference can personalize the response, while retrieval supplies evidence for it.

Row-Level Security (RLS)

A database policy layer that scopes every read/write to the requesting user.

RLS can restrict database rows according to policies for the requesting user. Its protection depends on the policies and access paths in use; review the security documentation for Brainbaby's architecture.

Signed URL

A time-limited, cryptographically signed link to a private storage object.

A signed URL grants access to a specific resource for a limited time. Treat it as sensitive while it is valid; expiration and visibility depend on the storage system and asset.

Speakable Schema

A schema.org property identifying page sections suitable for text-to-speech.

Support varies between services. Adding structured data does not guarantee that a page will be quoted or displayed in a particular search feature.

SSE (Server-Sent Events)

A one-way HTTP streaming protocol used to push tokens from the server to the browser as they're generated.

SSE lets a server send a sequence of updates to a browser over one HTTP connection. Streaming behavior depends on the feature and network connection.

SWE-bench Verified

A benchmark that scores an LLM on solving real GitHub issues end-to-end.

The benchmark reports results on a selected set of issues. Use the same benchmark version and test setup when comparing results.

System Prompt

Instructions supplied by the application to guide model behavior in a conversation.

A system prompt might tell an assistant how to format replies or when to ask for confirmation. A switch in system instructions can change the behavior without retraining the model.

Tau-Bench

A benchmark for agentic tool use in customer-service style tasks.

Tau-Bench measures whether an LLM can complete real workflows requiring API calls and multi-turn negotiation.

Temperature

A sampling parameter that controls output randomness.

Lower values typically make output less variable. Temperature alone cannot establish whether an answer is accurate.

Text-to-Image

Generating images from a text prompt.

Brainbaby offers image generation from a prompt and editing with a source image. Models, styles, output details and available credits vary by workflow and account.

Text-to-Video

Generating video clips directly from a written prompt.

Brainbaby offers text-to-video and image-to-video workflows. Available resolution, duration and credit cost depend on the selected model and settings; review the options shown before generating.

Tokens

Units a model processes. A token can be a whole word, a word fragment or punctuation.

Token counts vary by tokenizer, language and input. A word count is only a rough proxy for a model’s context or usage.

Tool Use / Function Calling

A model’s ability to request an application function with structured arguments.

The model might request a calendar lookup with a date. The host application validates the arguments and decides whether it can run the tool.

Uncensored AI

An AI configuration that lifts default content filters for adult creative work.

Brainbaby's uncensored routing keeps age-gated adult prompts on dedicated models so the mainline models stay safe for work.

Vector Database

A database optimized for storing and searching high-dimensional embeddings.

A vector database can retrieve similar documents for a RAG workflow. Other workflows use keyword search or combine both approaches.

Vision (Multimodal) Model

An LLM that can also read images.

Brainbaby's mobile and web chat accept image uploads for OCR, chart reading, product identification, and visual QA.

Voice Mode

Real-time spoken conversation with an LLM.

Voice features can combine speech recognition, model responses and speech synthesis. Language availability and performance depend on the product and environment.

Zero-Data-Retention

A processing arrangement that limits whether request data is retained after use.

Zero retention is different from a promise not to sell data or train models. Brainbaby saves account and workspace data, and providers process requests under their applicable terms. Review the privacy and deletion policies for the current details.

Try a free browser tool

Convert a file or download a worksheet without creating an account.

See free tools and resources