instibrain
instibrain  /  glossary

One word for each thing, defined once.

The site, the docs and the code use the same words. Here are the ones that name a part of instibrain, then the AI words a buyer meets beside them, in eleven groups, each with a link to the page where it does its work.

1 · the store

What a brain is made of.

The brain, its spaces and pages, who reads them, and what they said before.

brain

Every space a person reads, as one whole: their Personal and Sealed, the team spaces of their account, and any space opened for a few people. It is served to the sync client over HTTP and, for its open spaces, to AIs over MCP.

on /brain
space

A set of markdown pages in folders, shared whole with one set of people: no folder inside it is shared on its own. Pages only some people should read go in a space of their own. A space is open or sealed, and belongs to a person or to the account.

on /brain
personal space

A space a person owns, born with a grant to them alone; its grants, not its kind, decide who reads it. Every person has two from their first sign-in: Personal, open, and Sealed, sealed.

on /spaces/personal
team space

A space the account owns, which only an owner or admin opens. Its managers read its history, and the account is its content owner, so an owner or admin names who else recovers it.

on /spaces/team
page

One unit of memory: a path, a body in markdown, and a version vector. The first heading is its title.

on /brain
folder

A path inside a space. A folder exists through the pages in it, and it is shared as its space is.

on /brain
reader

The read-only role, the first of three: a reader reads a space, an editor also writes it, and a manager also decides who it is shared with. Each person, group, or agent a space is shared with holds one, and it covers every page.

on /brain
label

How far a page's text may travel: public, internal, confidential, or restricted. A space has one, and a page can narrow it for itself, never widen it.

on /guides/instibrain-with-claude-code/ask-your-brain
held write

A write an AI makes in the hour after it reads a confidential or restricted page, into a space where that text would go wider. It waits until a manager of that space reads it and lands it.

on /why
the guard

The check on every save, run with no model: it refuses a secret in any space, and a card number, an IBAN, or a passport number in an open space. It matches patterns, so a secret of another shape passes.

on /security
recovery grant

A person's right to read what a space said before: old versions, the version list, removed pages, and the changes feed, under a path. A manager of the space holds it for the whole space, by the role; the content owner names anyone else who holds it, and it lasts only while that person reads the space.

on /spaces
review

Where a restored page waits, read by its person alone, until they check who reads the space and publish it as the page's next version. No restore goes live without one.

on /spaces
cockpit

Where a person sees the brain, at my.instibrain.ai: its spaces and who each is shared with, its specialists, every page, what waits for them, the log of which AI or agent wrote what, and the read trail.

on /markdown-editor
read trail

Every page read of a space, kept 30 days: who read it, through which computer, AI, agent, or the cockpit, and when. The space's managers see every row and anyone else their own; no AI reads it. One caller opening many pages within ten minutes is mailed to the managers.

on /security
inbox

The page Personal starts with, inbox.md. A line or half a thought goes in from any surface, and leaves only by becoming a page of its own or by deletion.

on /spaces/personal
surface

Anywhere a person or an AI reaches a brain from: a computer running the sync client, Claude on the web or on a phone.

on the landing page
2 · people and accounts

Who signs in, and who runs what.

An account holds its people, and an account role runs it without opening a page.

account

Who owns the team spaces: a company, a team, or one person. Its people, and the people it invites, see the display name its owner sets in Settings.

on /spaces/team
identity

Someone who signs in: one email and one password. One identity can be in several accounts, and the cockpit switches between them.

on /get-started
person

An identity in one account, with the name, role, and picture that account sees. The brain knows a person by this id, so a space can be shared with someone invited before they join.

on /spaces/team
account role

Owner, admin, or member. Owners and admins run the account: its people, its groups, and its team spaces. An account role opens no page.

on /spaces/team
administrator

Our staff, apart from customers: a sign-in of their own on the internal cockpit, management.instibrain.ai. An administrator's login opens no account and no brain, and every change they make there stays on the staff trail for a year.

on /privacy
paused account

An account our staff stopped for now. Nobody signs in to it and its devices and agents are refused, but nothing is ended.

on /terms
invite

A one-time link to join an account, valid seven days, mailed or handed over. A new address chooses a password with it; an address that signs in already joins with the one it has.

on /get-started
group, and Everyone

People of one account under one name; a space or a specialist shared with it reaches them all. Everyone is the group every account has: it holds every person of the account.

on /spaces
3 · structure

How a space keeps its shape.

The files that say how to work in a space, the tools that hold it to them, and the scripts it carries.

template

The files a space starts with. A new space starts from empty: its index.md, which opens with its purpose. Each person's Personal and Sealed start from their own templates.

on /spaces
skeleton

The few files that say what a space is and how to work in it: its index.md, the page of its working rules where it has one, and the index file of each top-level folder. Only a manager of the space changes, moves, or removes one.

on /for-agents
index file

The file that says what a folder holds and how to add to it: index.md, in every folder. Its first sentence is the folder's purpose. A space's own index.md opens with what the space is for.

on /spaces
structure tools

The MCP tools that give an AI a space's map and skeleton and keep new folders structured: space_map, read_skeleton, check_structure, create_folder, and move_folder. On a computer, instibrain check runs the same rules for any space, sealed ones included.

on /for-agents
kit name

A page or folder name that starts with insti but not with instibrain. Such names are the system's, so every write door refuses a new one inside a space.

on /for-agents
operation

A task with an end, days to weeks long: a folder under operations/ in a space, whose brief.md holds the goal. In an open space, its people and their AIs talk in a room of its own until it closes.

on /guides/instibrain-with-claude-code/an-operation
tool page

The page in tools/<name>/ that says how an AI reaches one system: the access it needs, the calls it makes, and the scripts it may run. The credential stays in your own vault.

on /works-with
access map

A page that says which systems your company runs, who owns each, and where each credential lives. It grants no access: each system still decides who opens what.

on /works-with
tool folder

tools/<name>/ in a space: a tool's page and the files it runs, synced to everyone the space is shared with. Only text files of a script's or a data file's type sync there, and a computer never writes one executable.

on /faq
approval

A manager's word for each file of a tool, by the SHA-256 of its bytes, in tools/<name>/approval.json. A computer runs or exports a file only at its approved hash, and in a sealed space only an approval a chain admin signed counts.

on /faq
4 · open and sealed, and the keys

What our server can read.

A space is open or sealed, and the keys are what a sealed space is made of.

open space

A space our server reads. Its pages are stored as markdown, so an AI holding no key, on the web or a phone, can read them. Every person's Personal is one.

on /spaces
sealed space

A space our server cannot read. A page is encrypted on your computer before it leaves, path included, and only the computers of everyone it is shared with, which hold the space's key, open it. No web AI reads it. Every person's Sealed is one.

on /spaces
area

The pages under one key, or under none: always one whole space. An open space is one open area, which our server reads; a sealed space is one encrypted area, which it cannot.

on /spaces
key_id

The name of the key a page is encrypted under, or none for a page of an open space. A name for a key, never a key.

on /how-it-works
area key

A random 256-bit key, one per sealed space and epoch, wrapped once for each device admitted to it.

on /security
page key

The key one page is encrypted under, derived from the area key, the page's id and the epoch. No page has a secret of its own.

on /security
device key

The three keypairs a device makes for itself and never sends: X25519 with ML-KEM-768 to receive wraps, and Ed25519 to sign chain entries. They sit in the device's keys.json.

on /security
wrap

A key encrypted for one device's public key. The server stores wraps of area keys and cannot open any of them.

on /security
epoch

One generation of an area key. Removing a device or admitting one starts the next; an admitted device gets no earlier one, and the same sync sends every live page up again under the new key.

on /security
device approval

A person's yes to a new computer's fingerprint before it gets a sealed space's key. A teammate's first computer is approved on a manager's computer, each later one on a computer of theirs already in. A no is kept, and that device gets nothing.

on /security
recovery bundle

One file holding the whole keystore, device keys and every area key, sealed under a passphrase with Argon2id. You keep it off the machine it came from, and a key taken since makes it stale.

on /faq
membership chain

The append-only, signed log of which devices hold a sealed space's area key. A manager's device signs each entry, which carries the key wrapped for that device, and the sync client verifies the chain from its first entry.

on /security
the Blind Vault

The name of the encryption design behind sealed spaces. It names the design and nothing else, never a place a page is put.

on /security
5 · sync

Between a page and a file.

What the sync client does on your computer, and what it leaves behind.

sync client

The part of instibrain that runs on your computer. It adopts, materializes, syncs and captures conflicts, and it holds the keys. Its command is instibrain, and each space it syncs is a folder under ~/instibrain/spaces/.

on /device
device

Any place you are: one of your computers, or one of your AIs. Each is you, never above your role, and the cockpit's Devices page lists them. In the encryption, a device is one folder a computer links, with keys of its own: it signs every request, and the brain holds only the public key.

on /computers
adopt

What the sync client does with a markdown file that has no id yet: it asks the brain for a new page id, writes the short header into the file, and pushes it.

on /device
materialize

Turning pulled pages into local files. The sync client materializes every page of the spaces it syncs as plain markdown.

on /device
cursor

Where a pull left off: seq, the space's own number, which each page takes when its write commits. A pull asks for every page numbered after its cursor, in commit order.

on /how-it-works
tombstone

What a deleted page leaves behind. The row stays and its content goes, so the next pull tells every computer the page was deleted.

on /how-it-works
conflict file

When a push returns 409, the sync client writes the server's version beside yours as a .conflict.md file and overwrites nothing. You merge the two, then run instibrain resolve.

on /device
6 · the rule

How the server decides a sync write.

The server's only judgement on a sync write, and the vocabulary it uses.

version vector

One counter per writer of a page: each synced folder that wrote it, and the brain, which counts every write through its own doors as one writer. Comparing two vectors tells the server whether a write has seen everything the stored copy has.

on /how-it-works
bump

Adding one to your own device's counter in a version vector, which records one local edit.

on /how-it-works
join

The least upper bound of two version vectors, counter by counter. The vector a merged page carries.

on /how-it-works
dominates

Vector a dominates b when every counter in a is at least b's and one is strictly greater. The one rule the server applies to a sync write.

on /how-it-works
409

The answer to a sync write whose version vector does not dominate the stored one. It carries the server's vector, so your device can merge. The server never merges a page itself.

on /how-it-works
7 · your AI and agents

Your own AI, and the team's agents.

How your AI connects and what it reaches, and what an agent is.

your AI

An AI that acts as you, wherever it runs: Claude Code, Codex, Claude on the web or the phone, or a program of yours on the API. It reads its reach, never more than you. Several can read one brain at once. It is never ours, and sync never waits on it.

on /agent
agent

A member of the team that is not a person, such as a support bot or a writer. The account owns it, an owner or an admin makes it, and a space is shared with it as with a person, at its own role.

on /spaces
room

Where your AIs, your agents, and your people talk, beside the spaces: each window, each AI login, and each agent is a voice of its own, and you read every room you are in. A message lasts 30 days, unless someone keeps it as a page.

on /rooms
window

One named Claude Code window on a computer. Its messages read as you, in that window, and it hears its rooms with nobody typing into it.

on /rooms
connection

One login of one of your AIs, approved once in the browser, with a credential of its own. You see it on Devices and sign it out there, which ends that login.

on /agent
reach

The spaces and specialists one of your AIs reads, each to read, or to read and write, never above your role. A new AI reaches nothing until you give it a place, on Devices.

on /security
proposal

A change an AI may not make itself, left in a space: a title, a body, and the page it is about. It belongs to the space: every editor and manager of the space sees it in the cockpit, and whoever makes the change closes it.

on /for-agents
specialist

A persona and a method an AI works as, shared with people, groups, and agents of its own, beside the spaces: ~/instibrain/specialists/<name>/ on a computer, made and shared in the cockpit. What a space knows for it is knowledge/<name>/ in that space.

on /specialists
the system

What defines instibrain, the same for everyone: the identity, the map, the staff in staff/, and the tools in tools/, each tool's how-to in its own folder. It is no space and is read-only; on a computer it is ~/instibrain/system/.

on /for-agents
MCP

The Model Context Protocol. The brain speaks it over Streamable HTTP, so an MCP client reads the open spaces of a brain with no SDK or plugin. Claude Code, Codex, and Claude have run on ours.

on /agent
MCP server

A program that offers an AI tools and data over MCP, as a connector does. The brain is one, at an address your AI adds once.

on /agent
brain.read, brain.write, brain.manage

The OAuth scopes a connected AI can hold. The approval page sets reading, and ticks writing and managing where the AI asked for them; the spaces it reaches are set on Devices. Managing acts as the AI's person, never above them: sharing, groups, invites, roles, removing a person, a new space, and the person's own computers, each change to people, access, or computers made only after the person's yes on the cockpit's Home. A fourth, brain.sync, is the sync client's; no AI that signs in through the browser holds it.

on /tools
8 · the market's words

What the market calls it.

The names a company meets when it puts AI to work, where a brain stands on each, the kinds of memory, and the line between your pages and your files.

context engineering

Choosing what an AI holds in mind before it answers: who it is, what is true, and how the work is done, loaded ahead of your first word. A brain is context engineering kept as pages your team owns and corrects.

on /context-engineering
progressive disclosure

How a brain loads into an AI: a stable identity first, an index of what exists, then one page at a time as the task needs it. Your AI knows who it is and reads only what the task needs.

on /for-agents
agent memory

What an AI carries from one session to the next. In a brain it is pages a person can open and fix, never a hidden store of vectors, and every AI you connect reads the same ones.

on /brain
semantic memory

What an AI holds true about its world: facts someone owns and would correct. In a brain, it is a page for each subject.

on /context-engineering
procedural memory

How an AI does each kind of task: steps that start from a situation. In a brain, it is a specialist's method or a tool page.

on /context-engineering
episodic memory

What happened, and who said what, with a date. A brain keeps it out of the pages: the talk stays in a room for 30 days, and what it taught goes on the page of its subject.

on /context-engineering
environment

The world an AI works in and acts on: a company's files, records, tickets, and mail, reached through its own connectors. It is no part of the AI's memory; a page holds the map to it.

on /context-engineering
institutional memory, company brain

The plain names for what a company knows: its rules, its decisions, and why it made them. A brain keeps them as pages in team spaces, so they stay when a person leaves.

on /why
the context layer

The layer between your people's AIs and your systems: what each AI knows and may do. The model makers keep the weights; your company owns this layer, and it survives every model change.

on /why
context and data

The test for what goes in a brain. Context is what your company knows and has decided, and it is pages. Data is what your company has, files, records, tickets, and mail; it stays in your own systems and reaches your AI through their connectors.

on /why
AI governance

The rules for what each AI may read and do, and the record that they held. In a brain: who each space is shared with, each AI's reach, a label on each page, and the read trail.

on /security
access control

Deciding who may open what. A brain shares each space whole with named people, groups, and agents, and an AI reads no more than its person.

on /spaces
knowledge management

Keeping what a company knows current, findable, and free of contradictions. In a brain, each fact is one page, and the curator proposes what to cut.

on /guides/instibrain-with-claude-code/keep-the-brain-clean
version history

Every earlier state of a page, each naming its writer. Our server keeps an open space's every version for 30 days, then one a day for a year, then one a week; a sealed space's, none.

on /guides/instibrain-with-claude-code/every-change-kept
multi-agent orchestration

Several AIs splitting one job: who answers, who hands off, and who checks. In a brain they talk in rooms, and the work waits in the inbox and in operations.

on /rooms
integrations

The links between an AI and the systems a company runs: its file store, its CRM, its tickets. The links themselves are the vendors' connectors; a tool page in the brain says how your AI uses each.

on /works-with
tool use

An AI calling a function outside itself, such as a search or a script, and reading what it returns. The brain offers its own tools to your AI over MCP.

on /tools
connector

The vendor's own door to its system: you sign in through it once, and your AI reads and acts there. The access map says which connector opens which system.

on /guides/claude-basics/connectors
second brain

The notes one person keeps so they can use them later. A company's second brain is pages its team keeps, each space with its own people, read by every AI it uses.

on /second-brain
retrieval

Finding the right passage among many when a question comes, and handing it to the model. Pages know. Retrieval finds.

on /why
RAG

Retrieval-augmented generation: find the passages a question needs, put them in the prompt, and answer from them with a citation. It is a finder, not a memory: what your company knows is pages.

on /guides/rag
evals

A fixed set of real questions, scored the same way after every change, so you see whether the change helped. A brain's evals are planned.

on /guides/around-the-model/evals-and-the-golden-set
benchmark

A public set of test questions that ranks models against each other. It scores a model on its own questions, never on your company's.

on /guides/around-the-model/evals-and-the-golden-set
observability

A record of what each AI did: what it read, what it wrote, and when. A brain names the writer of every version and keeps every read of a space for 30 days.

on /security
AI platform

Where a model runs, and whom you pay for it: its maker, a cloud, or your own servers. Never instibrain: you bring your AI and pay its vendor.

on /why
inference

Running a model to get an answer, billed by the token. Your AI's vendor runs it; the brain runs no model and calls nobody.

on /guides/how-a-model-works/running-a-model
9 · the model's words

What a model is, and what it reads.

The words under every AI you use: the model, what it reads in one call, and how an answer goes wrong.

large language model, LLM

A program that writes text one token at a time, each chosen from everything it read before it. Claude is one.

on /guides/ai-basics/the-three-levers
frontier model

One of the most capable models of the day, from the few labs that train them. Every company rents the same ones.

on /why
open weights

A model you may download and run on your own machines. A closed model runs only at its maker, behind an API.

on /guides/ai-basics/fine-tuning-versus-context
weights

The billions of numbers inside a model, which training sets and nothing you type changes. The model makers keep the weights; the company owns the context layer.

on /guides/ai-basics/the-three-levers
knowledge cutoff

The date after which a model read nothing. What your company decided since reaches it only as something it reads, such as a page.

on /guides/ai-basics/fine-tuning-versus-context
fine-tuning

Further training of a finished model on your examples. It teaches a way of answering more than facts, and a fact it learned names no source.

on /guides/ai-basics/fine-tuning-versus-context
LoRA

Low-rank adaptation: fine-tuning that trains a small add-on beside a model and leaves its weights as they were. The add-on is small, so one model can carry one per task.

on /guides/ai-basics/fine-tuning-versus-context
transformer

The design under nearly every model that reads or writes language since 2017: a table of token vectors, a stack of layers of attention and feed-forward, and a last step that picks the next token.

on /guides/how-a-model-works/kinds-of-models
attention

The step inside a model where each token takes a share of the tokens before it, so a word gets its meaning in this sentence. The shares come from dot products and add up to 1.

on /guides/how-a-model-works/attention
temperature

How boldly a model picks its next token: at 0 always the likeliest, higher takes less likely ones more often. All the chance in an answer comes from this step.

on /guides/how-a-model-works/picking-the-next-token
quantization

Storing each of a model's weights in fewer bits, 8 or 4 in place of 16, so the model needs a half or a quarter of the memory, with a small loss of quality.

on /guides/how-a-model-works/running-a-model
reasoning model

A chat model trained to write its working before it answers, in thinking tokens you pay for. Higher effort in Claude Code means more of them.

on /guides/how-a-model-works/training
token

The unit a model reads, writes, and bills by: about three-quarters of a word, so 1,000 tokens are about 750 words.

on /guides/ai-basics/tokens-and-the-context-window
context window

How many tokens a model takes in one call, what it reads and what it writes together: one million on current Claude models.

on /guides/ai-basics/tokens-and-the-context-window
prompt

Everything a model is sent in one call: its standing instructions, what it read, and your request.

on /guides/rag/the-prompt-and-the-citation-rule
system prompt

The standing instructions at the top of a prompt, set by whoever runs the model: who it is and which rules it keeps.

on /guides/rag/the-prompt-and-the-citation-rule
prompt engineering

Writing one request so the model answers it well. Context engineering is the larger job: everything the model reads before that request.

on /guides/claude-basics/context
instruction file

A text file an AI reads at the start of every session: CLAUDE.md for Claude Code, AGENTS.md for Codex. On a linked computer the sync client builds both in ~/instibrain, from the identity and your profile, so every AI starts from the same brain.

on /guides/claude-basics/md-files-and-claude-md
skill

A method in a folder, which an AI loads only when a task needs it: procedural memory, with no facts of its own. In a brain, a method a team shares is a specialist.

on /specialists
prompt caching

The vendor keeps the start of a prompt it has read, so a call that begins the same way pays a fraction for it. An instruction file read at every start costs little for that reason.

on /guides/ai-hygiene/the-stable-part-first
context rot

Anthropic's name for a model's accuracy and recall going down as its window fills. A new window for a new subject keeps it away.

on /guides/ai-hygiene/one-subject-one-window
compaction

Claude Code replacing a long conversation with a short summary of it when the context window fills. A fact survives it whole only when it is on a page.

on /guides/claude-basics/context
hallucination

A fluent answer that is false: the model filled a gap with what sounds right. Grounding is the cure.

on /guides/rag/the-prompt-and-the-citation-rule
grounding, citation

An answer tied to the passage it came from, which it names so a reader can check it. A page grounds an answer, and the model can still misread it.

on /guides/ai-basics/rag-one-question-end-to-end
prompt injection

Text in a page, a file, or a mail that an AI obeys as if its person had asked. Whoever writes a space you read can steer your AI, so a write after a narrow read waits as a held write.

on /security
guardrails

Checks in code around an AI: what it may read, what it may write, and what it may send. A brain's are the guard and the held write.

on /security
human in the loop

A person's yes before an AI acts: a send, a delete, a share. An AI that manages a brain asks its person before each change to people, access, or computers.

on /guides/claude-basics/permissions
lost in the middle

A model attending most to the start and the end of its context, and least to the middle. So the instructions go first and the question last.

on /guides/around-the-model/the-context-window
working memory

The context window seen as memory: what a model holds for one call, gone after it. What must last goes on a page.

on /guides/around-the-model/the-kinds-of-memory
long context

Pasting everything into a very large context window in place of choosing. It breaks who may read what, dilutes every line, and is old after the first edit.

on /guides/around-the-model/the-stack-in-these-words
instruction hierarchy

A model trained to weigh the app's instructions above the person's message, and the message above a tool's result. Training makes that likelier, never sure.

on /guides/around-the-model/priming-and-sycophancy
priming

What is already in the context tilting what comes next. The pages that load first set the tone of every answer, so they are written calm and in positive form.

on /guides/around-the-model/priming-and-sycophancy
sycophancy

A model bending the facts to please: it drops a right answer when pushed, with no new fact. Priming tilts the tone, sentiment is the tone, and sycophancy bends the facts to it.

on /guides/around-the-model/priming-and-sycophancy
10 · around the model

What people build around a model.

The loop that lets a model act, the attack on it and what stops it, and the test that says it works.

agent loop

Model, tool, result, model again, until the model writes text. A program runs the loop and the model only writes; the market calls the whole an AI agent.

on /guides/around-the-model/tools-and-the-agent-loop
subagent

A separate call with a context window of its own, given one task, that returns only its answer. On a linked computer, each specialist is a Claude Code subagent.

on /guides/around-the-model/subagents-and-orchestration
workflow, agentic

Two ways to order the steps: a workflow fixes them in code, and agentic lets the model choose each next one. Most systems mix the two.

on /guides/around-the-model/subagents-and-orchestration
sandbox

A closed place where an AI's commands run without reaching the rest of the computer.

on /guides/around-the-model/subagents-and-orchestration
jailbreak

A prompt injection the person types in themselves, to talk an AI out of its rules.

on /guides/around-the-model/prompt-injection
tool poisoning

Orders hidden in a tool server's description of its own tools, which an AI reads when it connects.

on /guides/around-the-model/prompt-injection
memory poisoning

A prompt injection that gets itself written into a note or a page and fires in every session after. A page's versions name each writer, so a person finds the line and who wrote it.

on /guides/around-the-model/prompt-injection
exfiltration

An injection making an AI send data out, often hidden in a link or an image address.

on /guides/around-the-model/prompt-injection
lethal trifecta

Private data, text from outside, and a way to send data out, in one AI's reach. Together they let an injection steal; take one away and that path closes.

on /guides/around-the-model/prompt-injection
least privilege

An AI reaching only what its task needs. Your AI starts with no space, and reaches only those you give it.

on /guides/around-the-model/what-works-against-injection
taint tracking

A read of private data marking the session, so a later write to a wider place waits for a person. In a brain it is the held write.

on /guides/around-the-model/what-works-against-injection
defence in depth

Layers that each stop what the one before missed, since no AI that uses tools is fully safe.

on /guides/around-the-model/what-works-against-injection
red teaming

Attacking your own system on purpose, to find a hole before someone else does.

on /guides/around-the-model/what-works-against-injection
golden set

Real questions, each with its right answer or none, marked by people and run again after every change. Every eval starts from one.

on /guides/around-the-model/evals-and-the-golden-set
abstention

An AI saying it does not know. Right when the answer is not in its text; a false abstention when it is.

on /guides/around-the-model/evals-and-the-golden-set
live judge, LLM-as-judge

A second model that reads each answer beside the text it rests on, and stops an answer the text does not support.

on /guides/around-the-model/evals-and-the-golden-set
11 · how a search finds

How a search finds a passage.

What a file becomes before a search can find it, and how the search ranks what it found.

embedding

A list of numbers a small model writes for a text, so texts that mean the same get numbers close together: 768 to 3,072 of them with today's models.

on /guides/ai-basics/an-embedding-is-a-vector-made-from-text
vector

A list of numbers, compared with another by the angle between them. An embedding is one.

on /guides/ai-basics/a-vector-is-a-list-of-numbers
vector database

Software that stores vectors by the million and finds those nearest a question's. In a brain the pages stay the memory, and vectors only point to them.

on /guides/ai-basics/chunking-and-the-index
chunk

One slice of a document, often one section, which a search embeds, stores, and returns. Where the cuts fall decides what a search can find.

on /guides/rag/chunking-by-structure
index, ingestion

The table of chunks, their vectors, and their sources that a search runs against, and the job that fills it. Run the job again when a file changes, or the search returns old text.

on /guides/rag/ingestion-as-a-job
reranker

A second model that reads the question beside each passage found and puts the best first. Only the top few reach the prompt.

on /guides/rag/the-reranker

A word missing? Tell us, or read the FAQ.

Put these words to work on your own company's pages.

Ask for an account for your team, or take the tour first.