General AI
| Term | Definition |
|---|---|
| Prompt | The text you send to a model: a question, a statement, or a request. |
| Reply | The model’s response to a prompt. Also called a completion. |
| Token | A small piece of text a model reads or writes. A word may be one token or several. Speeds in Chat often use tok/s (tokens per second). |
| Context window | How much text (in tokens) the model can consider at once, including the prompt, recent messages, and the reply so far. |
| Prefill speed | How fast the model processed the prompt, in tok/s. Shown under Response stats. |
| Prompt eval | Prompt tokens that needed evaluation this turn (total prompt tokens minus cached). Shown under Response stats. |
| Cached prompt | Prompt tokens reused from the session cache. Shown under Response stats. |
| Decode speed | How fast the model wrote the reply, in tok/s. Shown under Response stats. |
| Generated tokens | Tokens in the answer. Shown under Response stats. |
| First token | Time to the first generated token after prefill. It does not include prompt-eval wall time. Shown under Response stats. |
| Response time | Prefill plus generation wall time for that turn. Shown under Response stats. |
| Context usage | Tokens used this turn over the model’s Context window, shown as used / total with a percent. See Response stats. |
| Parameters | The learned weights inside a model. A label like “7B parameters” describes size, not speed. |
| Inference | Running a model to produce a reply from a prompt. |
| System prompt | Instructions that shape tone, role, and limits before the user prompt. In webAI, Personas and Skills carry this kind of instruction. |
| Temperature | A setting that makes replies more fixed (lower) or more varied (higher). |
| Model | The trained program that turns prompts into replies. |
| Local model | A model that runs on your computer when you choose that path. A cloud model runs on someone else’s servers. |
| Quantization | A smaller, less precise copy of a model that needs less memory. Model settings may ask you to pick a quantization. See Model. |
| Fine-tuning | Further training a model on extra examples so it fits a narrower task. |
| LoRA | A small adapter loaded on top of a base model to specialize behavior without replacing the full model. |
| Embedding | A numeric representation of text that captures meaning so similar items sit close together. |
| Embedding model | A model built to turn text into embeddings for search and comparison. |
| Chunk | A short piece of a longer document, split so search and retrieval can work on manageable units. |
| Chunking | The process of splitting documents into chunks. |
| Vector search | Finding text by similarity of embeddings, not only by exact keywords. |
| RAG | Retrieval-augmented generation: fetch relevant chunks first, then ask the model to answer with that material. |
| Hallucination | A confident reply that is wrong or invented. |
| Grounding | Anchoring a reply in supplied documents, tools, or facts instead of relying only on the model’s prior training. |
| Tool calling | Letting the model choose an external action, such as searching Slack, when a prompt needs it. |
| MCP | Model Context Protocol, a shared format for exposing tools to a model. See Online & Connectors. |
webAI concepts
| Term | Definition |
|---|---|
| Collaborative intelligence | Specialized AI working with people, not only a single general model in the cloud. In the app that shows up as Personas, Spaces, Chat, Artifacts, and Skills. |
| Intelligence Delivery Network (IDN) | The connective layer behind peer-to-peer work in webAI. It is the network where intelligence runs, connects, and improves on hardware you control. Read more on the IDN platform page. |
| Peer-to-peer | Devices connect to each other for collaboration. In Spaces and Contacts, a message, a file, and a Persona call move between those devices. |
Surfaces
| Term | Definition |
|---|---|
| Chat | Where you send a prompt and use the reply. Threads stay under History. See Chat. |
| Auto | In Chat, lets webAI choose who answers instead of locking a Persona. |
| Answer as | On a reply, asks another Persona to take the same prompt. |
| Persona | A setup with a name, instructions, and tools you can chat with. See Personas. |
| Connected Persona | A Persona shared by someone else. It still runs on the device of the person who shared it. |
| Space | A shared place where invited people collaborate from their own devices. See Spaces. |
| Channel | A conversation inside a Space. Messages copy onto members’ devices. |
| Ask anything | A question you ask with the Space, not a private chat and not a channel message. |
| Shared knowledge | Documents added in a Space that people collaborating there can search. |
| Admin / Member | Roles in a Space. Admins invite people and manage the Space. |
| Artifact | A document a chat produced after you choose Yes, show dynamic doc. See Artifacts. |
| Skill | Saved instructions you apply to a prompt, often with / in Chat, or from a custom Persona’s Skills list. See Skills. |
| Contacts | People you add so you can invite them to a Space. See Contacts. |
| Generate invite / Pair | Create a one-time invite, then complete pairing so the other person appears in Contacts. |
Identity and network
Models and device
| Term | Definition |
|---|---|
| Downloaded model | A local model you apply in Settings so Chat and Personas can use it on this computer. See Model. |
| Thinking mode | When the selected model supports it, shows Thinking and Thought process in Chat. |
| Device tier | A read-only label from this computer’s memory. It sets which features start on at setup. See System. |
| Long-Term Memory | Facts about you that a chat on this device can use in a later reply. See Personalization. |