Conversations

This page covers the mechanics of working with the AI advisor — what is on the screen, what each control does, and what the response looks like as it streams in.

The composition area

The text input is anchored at the bottom of the screen and is composed of three controls:

• Depth dial — a small icon at the left toggles a popover with the three thinking-depth options (Surface, Balanced, Deep). The icon reflects the currently selected depth. The selection applies to the next message only; the default is restored after sending.
• Text area — supports multi-line input via Shift + Enter. Enter alone sends the message.
• Send button — disabled while a message is empty or a previous turn is still streaming.

Above the input, a small footer reads "AI can make mistakes. Check important info." — a standard reminder applicable to any AI-driven workflow.

The empty state

Before any conversation has been opened, the workspace shows a centred greeting ("Good morning, " followed by the user's display name) and the prompt "Can I help you with anything?". The composition area sits beneath the greeting and is positioned to mid-screen until the first message is sent. After the first send, the layout transitions: the messages area expands upward and the composition area docks at the bottom.

Mentioning files

Pressing @ in the input opens the file picker, a search dialog over the user's Repository and the public corpus.

• The dialog defaults to filename search — fast, exact, suited to "I know what the file is called".
• A toggle in the upper-right of the search box switches to content search — slower, semantic, suited to "I want files that are about X" even when the filename does not contain X.
• Selecting a result attaches it to the message as a file chip above the input. The chip is removable.
• A message may attach multiple files. The advisor sees the file as a first-class reference and is more likely to draw on it as a primary source.

Attaching a file is different from asking the advisor to search — it pre-loads the file directly, removing the search step. Attach a file when you know the answer is in it; let the advisor search when you don't.

The message timeline

Each turn produces a message timeline that streams in from the top of the response, before the prose. The timeline is the visible record of what the advisor did to produce its answer.

Reasoning blocks

Pieces of the advisor's working notes appear as collapsible rows headed "Thinking…" (while in progress) or "Working through the request" (once complete). They are open while the advisor is mid-thought and collapse automatically when the turn ends — the user can re-open any block to inspect the reasoning.

Reasoning blocks are not the final answer; they are the advisor's working notes. They are useful for understanding why a conclusion was reached, particularly when the conclusion is surprising.

Lookup chips

Every external lookup the advisor performs becomes a chip in the timeline. A chip carries:

• An icon identifying the area of the platform the lookup is against (voting records, MP profiles, public statements, alerts, documents, conversation continuity).
• A short label describing what the lookup did (e.g. List votings, Search documents).
• A state glyph: a spinner for running, a … for pending, an exclamation for error. Successful calls show their wall-clock duration (e.g. 1.2s).

Adjacent lookups are folded into a collapsible group ("Using N tools…" while running, "N tools used" afterwards) so a turn that performs eight lookups in sequence does not produce an eight-row block by default.

Elicitation forms

Some lookups require structured input from the user mid-turn — choosing one of several Sejm terms, picking the right MP from a name-collision, confirming a destructive action. When this happens, the lookup pauses and an inline form appears beneath the chip.

• Free-text inputs render as text fields.
• Single-choice values render as dropdowns.
• Boolean confirmations render as checkboxes.
• Object-shaped requests render as a stack of named fields.

The user submits or skips. The submitted value is recorded inline ("You answered: ..."); skipped requests fall back to a sensible default. Once resolved, the lookup resumes and produces its final result.

The advisor does not duplicate this behaviour in its own prose — when a lookup can disambiguate via its own form, the advisor uses it rather than asking the user the same question in text.

Citations

The body of an advisor response uses two distinct citation styles, both rendered specially.

Resource chips

References to records on the platform are rendered as inline chips with the source's icon, a label, and a click target. Selecting a chip opens the corresponding record:

• A voting chip opens the voting detail in Voting ledger.
• An MP chip opens the stakeholder profile.
• A post chip opens the original post on X.
• An alert chip opens the alert detail.
• A file chip opens the file in Repository (where applicable, scrolled to the cited passage).

The chips are stable across regenerations — re-running a question that produced a particular set of citations will normally produce the same chips again.

File citations

Numbered superscript chips inline with the prose are file citations. Hovering reveals the source filename with its file-format icon; clicking downloads the file.

This style is reserved for content search results over uploaded files — when the answer is grounded in a passage the user themselves uploaded, the citation surfaces the original document directly.

Per-message controls

Beneath each completed assistant message:

• Copy — copies the prose body to the clipboard. Citation markup is replaced with the underlying text representation.
• Regenerate — reissues the user's previous message and produces a new response. The original response is replaced; previous turns above are unaffected.

There is no edit-and-resend on user messages today.

Conversation history

The right column lists every conversation on the user's account.

• Advisor History is the column heading.
• New conversation at the top clears the workspace.
• Past conversations are listed by the topic the platform inferred from their first messages. New conversation is the placeholder before a topic is determined.
• Selecting a conversation reopens it. The transcript is reloaded with full fidelity — reasoning blocks, lookup chips, citation chips, and any unresolved elicitation forms are all reconstructed.
• The trash control beside each entry deletes the conversation and its transcript permanently. The control reveals on hover for inactive conversations and is always visible for the active one.

On narrow viewports, the history column is hidden behind a History button in the top-right of the workspace; selecting it slides up a bottom drawer with the same content.

Refresh and resume

A turn that is in progress when the user navigates away from the conversation, refreshes the browser, or loses connectivity for a moment is not lost. Reopening the conversation reattaches to the live response and replays whatever has happened so far before it resumes streaming.

In practice this means it is safe to:

• Send a Deep query and switch to Live proceedings to do something else while the advisor works.
• Refresh the browser deliberately mid-stream — the response continues to assemble.
• Open the conversation in a second tab — both tabs see the same stream.

The only failure mode that interrupts a turn is the user explicitly deleting the conversation while it is still streaming.

Errors

When something goes wrong during a turn, an error block is displayed beneath the partial response with a short description of what failed.

The most common causes are: the turn legitimately ran out of capacity for the question, a backing data source was temporarily unavailable, or the question is outside the scope the advisor is built to answer.

Regenerating a failed turn is a reasonable first response. If the failure persists, the question may benefit from being narrowed or split into parts.