Skip to main content
Version: Peki

Glossary

Every term you meet in Peki, from A to Z. UI labels are in bold. Follow the links for the full explanation.

A​

  • Action: The Function an automation runs to write to another system, without using a model. Its writes are approved once, when you save the automation. See Automations.
  • Admin: The role that manages members, API keys, secrets and workspace settings, on top of everything an Editor can do. See Members and roles.
  • Agent: One AI assistant with one job, for example "ShopX order support". It has its own instructions, tools, knowledge, channels, conversations and analytics. See Key concepts.
  • Agent access: A records table setting that decides whether the agent may read, create or update records in that table. Tables of people keep read off, so one customer can never ask about another. See Records.
  • Agent document: See Document.
  • Agent settings: The page with an agent's name, creation date, live version, model and the Delete agent action.
  • Agent team: Other agents of your workspace that an agent can ask for help while it answers its own customer. Each team member becomes a tool named after it. See Agent teams.
  • Allowed websites: The list of sites that may show your website chat. Empty means any site. See Website chat.
  • Analytics: The page that shows how well an agent answers and what it brings in, from live conversations of the published agent. See Analytics.
  • Answered: The evaluation's verdict on a finished conversation: whether the customer's question was answered (yes, partly or no).
  • API key: A workspace key, starting with psk_, for calling the Peki API from your own systems. It has Read only or Read and write access. Not to be confused with a Secret. See API keys.
  • Apply to the agent: A button on a Document section you edited. It asks the Copilot to change the agent to match. See The Agent document.
  • Approval: A request for your team to check an action before it runs: a tool call, a Function, a records save or an automation message. It waits in Inbox → Approvals and expires after 24 hours. See Approvals.
  • Approval card: A Copilot card that asks you to confirm an action with real effects, such as publishing, restoring a version or an automation that writes. You click Approve or Deny.
  • Approve and build: The button on the Copilot's Approve the Agent document card during onboarding. After it, the Copilot builds and tests the agent.
  • Automation: A job an agent runs on its own: on a schedule, from a webhook or on a Peki event. It can check with a Function, compare the result, run a task with the model, and deliver a Telegram message or other output. See Automations.

B​

  • Brief: What a main agent sends a team member when it asks for help: the goal, the facts the member's tools need and the answer format.
  • Bubble: The website chat style with a button in the corner of every page; the chat opens over the page.
  • Built-in variables: Variables Peki fills on every turn: customer.name, business.name and today. You can't create or change them. See Variables and secrets.
  • Business hours: Your weekly opening windows, in the workspace time zone. Used by conditional instruction blocks, the agent's sense of whether you're open, and the agent's date and time. See Business hours.

C​

  • Card: A place where the Copilot stops for you: options, an API key field, a Telegram bot token field, an approval, a new table, a Document section suggestion or the Document approval.
  • Channel: A place where customers reach an agent: Website chat, Telegram or the Sessions API. See Channels overview.
  • Chat box: The website chat style that shows an always-open chat box inside your page, where you paste the code.
  • Check: The first step of an automation: a Function that runs without the model and costs almost nothing, so the automation only continues when something happened.
  • Compact history: A Copilot option. The Copilot gets a summary of older messages and the last few in full, which makes long chats much cheaper.
  • Condition: A rule that decides whether something happens. An instruction block's condition (channel, business hours, customer tag or customer language) decides whether the block is sent. An automation's condition, like rsi14_1h < 30, decides whether it continues.
  • Conditional: An instruction block that is sent only when its condition is true. The opposite is Always active.
  • Context: Agent settings for what the agent knows besides the conversation: the Current date and time and the Customer profile.
  • Conversation: One exchange between a customer and an agent, with its messages, files, traces, handoff events and rating. Also called a session. See Inbox.
  • Copilot: The AI assistant inside Peki that builds and changes an agent with you. It edits the draft, tests with real calls and stops for your approval before anything with real effects. See Copilot overview.
  • Credits: What a workspace spends on everything that uses AI. One credit is $0.01. See Plans and credits.
  • Customer: A person who talks to an agent. Peki creates one automatically on their first message on a channel. See Customers.
  • Customer details: The Inbox side panel showing a customer's name, tags, notes, channel, handle, language, first seen, last message and conversations.
  • Customer profile: An agent setting. When it's on, the agent sees what's known about the customer, such as their name and tags.

D​

  • Default language: The workspace language, English or Turkish. The agent uses it for texts it writes before it knows the customer's language.
  • Delivery: One call another system made to an automation's webhook URL. Requests shows the last 20 with a masked preview.
  • Document: The agent's description in plain words: business, job, what it can do, what it takes in, what it gives back, limits, handoff, tone and languages, channels and open items. The Copilot keeps it current; sections you write are yours. See The Agent document.
  • Draft: The editable copy of an agent. You and the Copilot change it; the Test panel uses it. Publish turns it into the next version. See Publishing and versions.
  • Dry run: A test run of an automation that shows what would happen but sends and writes nothing.

E​

  • Early access: The sign-in mode while Peki opens by invitation. New people without an invitation join the waiting list. See Accounts and workspaces.
  • Editor: The role that builds agents: edits the draft, uses the Copilot, tests, publishes, sets up channels and automations.
  • Embed code: The one-line <script> tag you paste into your website to show the website chat.
  • Evaluation: A small model reading each finished live conversation, with personal data hidden, to decide whether the question was answered and give it a topic. It can run on every conversation, on 1 in 5, on 1 in 10, or be off.

F​

  • Fallback model: The model that takes over when the primary model fails or stops responding. Set it in Instructions → Model.
  • Fallback reply: The message sent when the agent can't answer at all, for example when every model is down. Leave it empty to send nothing.
  • Fill from the conversation: A records table option. A small model fills the table from a conversation 10 minutes after it goes quiet. Also called the records fill.
  • Free plan: The plan every new workspace starts on. You can build, test and connect channels; publishing needs a package.
  • Function: A small TypeScript program the agent can call as a tool, running in a locked sandbox that reaches only the sites you allow. See Functions.

G​

  • Guardrails: Checks on messages and replies: block prompt injection, stay on topic, never reveal instructions, block offensive messages, plus personal-data protection. See Guardrails and privacy.
  • Guide: The in-app explanation of how Peki works, under Guide in the sidebar.

H​

  • Hand back to AI: The Inbox action that returns a taken-over conversation to the agent.
  • Handoff: How a customer reaches a person. The conversation shows under Needs you in the Inbox and a teammate takes over. See Human handoff.
  • HTTP tool: One API call the agent can make, with its inputs, headers, key and the response fields it keeps. See Tools.

I​

  • Inbox: Every conversation of an agent in one place, updating live, with Needs you, Taken over, All, Approvals and Team calls. See Inbox.
  • Instructions: The ordered prompt blocks an agent always follows, plus the tabs for model, memory, context, fallback reply, guardrails, handoff, web search, team and versions. See Instructions.

K​

  • Knowledge: Documents, web pages and text the agent searches before it answers. Changes apply right away. See Knowledge.
  • Knowledge source: One file, web page or text in an agent's Knowledge.

L​

  • Leads: In Analytics, records created during conversations.
  • Live: The status of an agent with a published version, and the label of the version customers talk to.
  • Live conversation: A conversation with a real customer and the published agent. Only live conversations count in Analytics.

M​

  • Match on: Up to three fields of a records table (such as phone or email) that identify one person or thing, so a save with the same values updates that record instead of adding a new one.
  • Memory: How many earlier messages of a conversation the agent sees (default 20), and when a new session starts.
  • Model: The AI model that writes an agent's replies. Each agent has a primary model and an optional fallback. The Copilot has its own model picker. See Model and behavior.

N​

  • Needs approval: A setting on a tool, a Function or a table's saves. The action waits in Inbox → Approvals until a teammate decides.
  • Needs you: The Inbox list of conversations where the customer asked for a person and nobody has taken over yet.

O​

  • Onboarding: Building a new agent with the Copilot from your website: it reads the site, asks a few questions, writes the Document and builds and tests the agent. See New agent onboarding.
  • Operator: The role that answers customers in the Inbox, on top of everything a Viewer can do.
  • Owner: The role that can do everything, including managing other owners and deleting the workspace. Every workspace has at least one.

P​

  • Package: What a workspace needs to publish agents. Packages are coming soon.
  • Paid plan: A workspace plan that lets agents go live.
  • Peki event: Something that happens inside Peki and can start an automation: record.created, record.updated, conversation.handed_off, conversation.resolved or approval.decided. See Triggers.
  • Personal data masking: Turning names, phone numbers, emails, ID, card and IBAN numbers into placeholders like [PHONE_1] before the model sees them. Your tools still get the real values.
  • Placeholder: The token that stands in for a piece of personal data in what the model sees, for example [TCKN_1] or [NAME_1].
  • Prompt block: One titled block of instructions. Blocks are sent in order, always or only when their condition holds.
  • Publish: Turning the draft into the next numbered version and making it live on every connected channel. Needs a package.

Q​

  • Quiet hours: An automation limit. In quiet hours its messages are held; runs are still recorded.

R​

  • Records: Tables of data the agent collects and keeps, such as leads, appointments, orders and support requests. Not versioned. See Records.
  • Records fill: See Fill from the conversation.
  • Resolve: Closing a conversation in the Inbox or through the API. The customer's next message starts a new session.
  • Restore to draft: Copying an earlier version into the draft. Nothing goes live until you publish.
  • Role: What a member may do in a workspace: Viewer, Operator, Editor, Admin or Owner.
  • Run: One execution of an automation or a Function. Automation runs show on the Runs tab.

S​

  • Secret: An encrypted API key that an agent's tools or Functions use, tied to the sites it may be sent to. Only the last four characters are ever shown. See Variables and secrets.
  • Session: See Conversation.
  • Sessions API: The API for running chats in your own app: start sessions, send messages, stream replies. See API overview.
  • Signing secret: A secret that only checks the signatures of incoming webhook requests (for example from Shopify or Stripe) and is never sent anywhere.
  • Skill: How-to steps the agent loads only when a matching kind of question comes up, so its always-on instructions stay short. See Skills.
  • Starter questions: Up to four example questions the Test panel offers while its chat is empty.
  • Store: An automation's small memory between runs: up to 50 notes, so it only reports what's new.

T​

  • Table: One kind of record in Records, with its fields, title field, match-on fields and agent access.
  • Take over: The Inbox action that lets a teammate reply to the customer. The agent stays quiet until Hand back to AI.
  • Task: The automation step where the agent writes a message or does the work, using only the tools you approved for it.
  • Team call: One request from a main agent to a team member. Shows in the member's Inbox under Team calls.
  • Temperature: How varied the agent's replies are. Lower is more consistent, higher is more creative. Default 0.3.
  • Test conversation: A conversation from the Test panel. It uses the draft and never counts in Analytics.
  • Test panel: Where you chat with the draft as a customer would. Opened with Test.
  • Time saved: In Analytics, conversations resolved without a person multiplied by the minutes per conversation you set.
  • Tool: An action the agent can take during a conversation: an HTTP tool, a Function, a team member or a built-in tool like web search or handoff.
  • Topic: The subject the evaluation gives a conversation. Topics that often go unresolved show "A skill for this would help".
  • Trace: The record of one reply: the model, prompt blocks, skills, tools, Knowledge, tokens and timing. Shown under Why? in the Inbox.
  • Trigger: What starts an automation: a schedule, a webhook or a Peki event. See Triggers.
  • Turn: One Copilot reply to your message, with all the steps it took. Each turn can be undone.

U​

  • Undo this turn: Reverting everything one Copilot turn changed in the draft. If you changed the same things since, it lists what would be lost and asks first.
  • Usage: The settings page showing what used your credits, by kind, agent and day. See Usage.

V​

  • Variable: A workspace value you use as {{name}} in prompt blocks. Changing it applies to every agent right away, without publishing.
  • Version: A published, numbered snapshot of an agent: v1, v2 and so on. One version is live at a time.
  • Viewer: The read-only role.

W​

  • Waiting list: Where new people's sign-up requests wait during early access, until access is ready.
  • Web search: An agent setting that lets it search the web for current information. Off by default.
  • Webhook: A URL another system calls to start an automation the moment something happens there. See Triggers.
  • Website chat: The chat on your website, as a bubble or a chat box, added with the embed code. See Website chat.
  • Widget tester: The customer name of conversations you start with Try it on the Channels page. They are real conversations with the published agent.
  • Workspace: A separate space with its own agents, channels, team, keys and credits. See Accounts and workspaces.