Zum Inhalt springen

Pia assistant

Dieser Inhalt ist noch nicht in deiner Sprache verfügbar.

Pia is a guide built into this console. It answers questions about Pia Server from the bundled server documentation and the release notes, and it knows which admin page you are on (the page’s route, title and area travel with each message; see What leaves the server). It is read-only: it explains and links, it never changes anything. You make every change yourself.

Pia answers in the language you write in. The bundled documentation is English only; Pia translates it into your language when it answers. Links in its answers open admin pages you can access, or pages on docs.pia-ai.de, which are marked External website and open in a new tab.

Every admin page has a round P in the bottom-right corner. It is semi-transparent until you hover over it or open it. A small dot on the bubble means Pia has a message waiting for you; a screen reader announces New message from Pia while the panel is closed, and the bubble’s name changes to Open Pia assistant, new message. Open the panel, type a question in the Message to Pia box and press Enter (Shift+Enter starts a new line) or Send.

While a dialog is open the bubble and the panel are hidden, and they come back when the dialog closes.

While the console shows its reconnecting dialog, the message box and the attach and send buttons are switched off, so nothing you type can be lost. When the connection is back they work again and whatever is in the box is kept.

Pia writes first in three situations:

  • A new admin gets a welcome with a short tour of the main areas and the setup steps that are still open. The server decides which steps are open, so Pia never lists something that is already done.
  • An admin who used the console before the assistant existed gets an introduction to Pia and a summary of the release notes for this version.
  • After an update, Pia writes a briefing that covers every version since the admin’s last one, with the Action required items first.

Each message appears once per admin and version. Signing in again shows nothing new, and if nothing has changed since your last briefing, Pia stays quiet.

Pia’s database access is controlled by a switch in Settings, on the Pia assistant tab: Assistant database access. It is off by default.

  • Off: Pia has only the documentation, the release notes and the page you are on.
  • On: Pia can also read this server’s configuration: groups, AI providers and where each model is used, knowledge bases, managed personas, connectors and operators, plugins, client policies, limits, the licence status and the server version.

Pia never reads secrets: API keys, tokens, password hashes, encrypted keys and connection strings are not readable by any of its lookups.

Pia reads only the areas your admin access lets you open. An admin whose bundle grants only Groups gets nothing about AI providers or licensing from Pia.

Questions about a person need your approval for every single lookup. With Assistant database access on, Pia can ask for these, each one only if your admin access lets you open the area it belongs to:

Lookup What Pia gets Area
Account by e-mail address id, address, display name, company, sign-in provider, creation and last sign-in time, whether it is disabled, its group Accounts
Group of an account the group, its organisation and whether the person manages the organisation or its knowledge bases (the organisation and the manager flags only if you may open Groups) Accounts
Recent sign-ins the 20 latest attempts: time, outcome, method, country and the client family (browser and major version with the operating system, such as Chrome 130 / Windows) Accounts
Devices up to 50 devices: name, platform, app version, status and times Sync
Sign-in anomalies the 20 latest: kind, when it was detected and whether it was acknowledged Audit log
AI usage totals over 1 to 90 days (30 by default): requests, tokens, credits and the five models that used the most Token usage

Sign-ins carry no IP addresses, cities or coordinates, and no raw browser text. Devices come without ids or keys. AI usage is totals only, never prompts or answers. Lookups go by e-mail address; when two accounts share one, Pia asks which one you mean.

Before each lookup Pia puts a card in the panel, for example:

Pia wants to look up the account of anna@example.com. This sends personal data to the AI provider Text EU and is recorded in the audit log.

The first part changes with the lookup (look up the last 30 days of AI usage of, with the number of days Pia asked for, look up the group of, list the recent sign-ins of, list the devices of, list the sign-in anomalies of). The card shows the full e-mail address so you can see exactly whom the lookup is about, and names the AI provider that receives the answer. Choose Allow or Deny. While a card is open, the panel says Pia is waiting for your answer., and after you decide the card shows Allowed or Declined and keyboard focus moves to that line.

After Deny, Pia is told that the admin declined and answers without the data. A card that is still open when you close the browser tab or reload the page is not allowed; the card and the audit log record it as cancelled. Every lookup asks again: an earlier Allow never covers the next call.

Each decision is written to the audit log as AdminAssistant.ApprovalDecided with the tool, the decision (allowed, denied or cancelled), the e-mail address shortened to a…@example.com and the argument summary. The log keeps no full address; a keyed hash of the address, derived from Encryption:MasterKey, lets the same person be recognised across entries without being readable.

Pia finds documentation through a system knowledge base, Pia Server Docs, which the server fills from the bundled documentation at start-up and after an update. It is listed on Knowledge bases like any other knowledge base. If the embedding model (used for semantic search) is not available, Pia falls back to plain text search over the same pages.

Pia is charged to your group, like a client chat. The admin who asks is a Pia user in one group, and that group’s AI credit limits apply to Pia as well: every answer and every proactive briefing counts against them. If the admin has no explicit group, the default group is charged.

When the group’s credit is used up, Pia says so in the panel. For the weekly limits it also tells you when it can answer again, in your browser’s time zone. The hourly and daily limits roll with use rather than resetting at a set time, so for those, and when free-tier access is suspended on the server, the message gives no end time. Pia answers again as soon as the limit resets, frees up or is raised.

Usage appears on Token usage under the template admin-assistant. When a vision provider reads an image for Pia, that use is charged to the same group and shows under vision:describe_image.

What When Sent to
Your message, and the current page’s route (ids masked), title and area Every message The Assistant-mode AI provider
Documentation excerpts and release notes When Pia looks something up The Assistant-mode AI provider
Configuration (no secrets) Assistant database access is on The Assistant-mode AI provider
Passages from knowledge bases not marked sensitive Assistant database access is on The Assistant-mode AI provider
Earlier messages and Pia’s answers in this conversation (last 20) Every message The same provider; dropped when the provider changes, when the page is reloaded, when you choose Clear conversation, or when an image goes to a separate vision provider after an approved person lookup
Which basic setup steps are still open (licence, embedding model, groups, accounts), and with database access on also a licence that expires within 30 days The first-time welcome The Assistant-mode AI provider
Search terms Pia writes for a documentation search Whenever Pia searches and the docs knowledge base is in use, whatever the Assistant database access switch says The embedding provider (Embeddings settings)
Search terms Pia writes for a knowledge-base search Assistant database access is on The embedding provider (Embeddings settings)
A person’s data (see Approvals) Assistant database access is on and you allowed that lookup The Assistant-mode AI provider
An image You allowed that image on its approval card The Assistant-mode AI provider, or the vision provider configured for it

Data from approved person lookups never goes to a separate vision provider: while an image is read through one, Pia does not offer the person lookups, and an earlier conversation that contains such data is set aside for good. The image itself, and the question Pia writes about it, do go to that provider.

The Assistant-mode provider is the AI provider assigned to Assistant mode. It is resolved the same way as for any Assistant request in your group: the group’s own routing first, then the server-wide mode assignment on AI providers, then the server’s configured default (Ai:Modes, then Ai:Default). Changing the group routing or the provider assignment changes where Pia’s conversation goes.

After you have approved a person lookup, Pia searches the documentation with the plain text search only, which calls no embedding provider, and does not offer the knowledge-base search, until that conversation is dropped.

Pia never uses the provider’s own web search or agents, even when they are switched on for that provider, so nothing is searched on the web on your behalf.

Show Pia what you see: click Attach image in the panel, paste a screenshot into the message box, or drop a file onto the panel. Each image shows as a small preview above the message box, and Remove image takes it off again.

  • Only PNG and BMP are accepted. Pia recognises them by their content, not by the file name. A BMP is converted to PNG in your browser, and a PNG is re-encoded there too, so metadata stored in the file never leaves your browser and only the pixels go out.
  • An image can be at most 5 MB and 8192 pixels wide and high, and one message can carry at most 3 images. Pia names the problem in the panel when an image does not fit, for example Only PNG and BMP images can be attached., Images must be 5 MB or smaller. or You can attach up to 3 images per message.
  • The server checks every image again before anything is sent. If one fails, nothing from that message is sent: Pia says which image could not be read (Image 2 could not be read as PNG or BMP. Nothing was sent; remove it and send again.), and your text and the other images stay in the message box.

Every image gets its own approval card with a preview of that image and a sentence such as Pia wants to send image 1 of 3 (1920 × 1080 px) to the AI provider Text EU. If the chat model cannot read images itself, the card also names the vision provider that reads the image on its behalf. Choose Allow or Deny for each image. A denied image is not sent, and if you deny all of them, no image is sent and a message without text is not sent at all (No image was approved, so nothing was sent.). Pia does not call the model.

An approved image goes only to the provider named on its card, and only while Pia answers that message. An automatic retry of the same answer may send it again to the same provider. Later messages never resend it: Pia only knows that an image was shared earlier.

  • If the Assistant-mode AI provider can read images, it receives them directly.
  • Otherwise Pia uses the image reading delegate configured for that provider on AI providers, if there is one. That provider describes the image, and the answering provider only receives the description. If the delegate was changed while a card was open, Pia sends no image and says The vision provider changed while you were deciding, so no image was sent. Send the image again to see the new provider.
  • When the delegate reads an image, Pia cannot look up personal data in that answer: the person lookups are not offered while the delegate is involved. If the earlier conversation contains data from an approved person lookup, Pia discards it for good and says Pia has set aside the earlier conversation, because it contained personal data and this image goes to another provider; later messages start fresh. If the vision provider turns out not to be usable, no image goes out, and Pia says Pia has set aside the earlier conversation because it contained personal data; later messages start fresh. The question Pia writes to the vision provider is limited to 1000 characters.
  • If neither exists, Pia says No vision-capable provider is configured; please describe what you see. and asks you to describe the screen instead.
  • A vision provider reads at most the number of images set by the server’s vision image limit, AiProxy:VisionMaxImages (default 8); further images are dropped before any card is shown.

Every approved image is recorded in the audit log as AdminAssistant.ImageSent, with its size in bytes and pixels and the provider names, and never the image itself.

By default each admin can send 6 messages per minute and 60 per hour. Messages Pia starts itself do not count. When you go over, Pia shows Too many messages. Wait a minute and try again.

Both limits are editable on the Pia assistant tab in Settings, as Messages per minute (1–60) and Messages per hour (1–1000, at least the per-minute value). Changing them needs Manage access to Settings; with View access you can read the values only.

The audit log records the assistant’s activity:

Event When
AdminAssistant.ToolCalled Pia used a lookup. The entry names the tool, the argument names the tool declares and the outcome (any other key is only counted, under <undeclared>); text values are reduced to their length, e-mail addresses are replaced by a keyed pseudonym, ids and yes/no values are kept, and a number is stored as it is only when the tool declares that parameter numeric (such as a list limit)
AdminAssistant.SettingsChanged Someone changed the switch or the rate limits
AdminAssistant.ApprovalDecided You allowed, denied or left a personal-data card unanswered. The entry has the tool, the decision, the shortened address and the same argument summary as a tool call
AdminAssistant.ImageSent You allowed an image and it was sent. The entry has the size in bytes and pixels and the provider names, never the image

The conversation text itself is not written to the audit log.

Nothing about the conversation is stored on the server. Reloading the page, opening a new tab or signing out starts a new, empty conversation. The audit entries and the usage rows remain. When the Assistant-mode provider changes or is edited in place, Pia says Pia’s AI provider changed, so earlier messages in this conversation are no longer sent. and starts without the earlier messages. Clear conversation (the bin icon in the panel header, available while Pia is not answering) starts fresh: it empties the panel and drops the earlier messages, so nothing is sent along with your next one.

The panel says An AI provider is needed before Pia can answer. and, if your admin access allows it, links to Open AI Providers. Without that access it asks you to contact an administrator who manages AI Providers.