Best practices

Formulating questions using corpus terms

The Advisor is designed with specific skills — analyzing and searching Sejm proceedings, legislative processes, stakeholders, clubs, votes, recordings, Media and files. Questions formulated precisely using these terms generate better and more accurate responses.

• Less effective: "What is happening with energy policy?" — the scope remains undefined.
• More effective: "Summarize recent floor activity on the renewable energy bill, including the last vote and committee actions from the past week."

Not every question needs to be elaborate. Including at least one specific piece of information — a bill name, print number, stakeholder, club, date range, committee, or file — is sufficient.

Choosing the thinking depth

The depth control in the composer applies to a single message:

• Surface — lookups with an obvious single answer and short follow-ups to an answer already given.
• Balanced — everyday questions; the standard setting.
• Deep — multi-source briefings, comparisons and questions where completeness matters more than speed.

Because the depth is set per message, one conversation can mix them: a Deep question for the briefing, then Surface follow-ups on its details.

Attaching context instead of describing it

When the answer resides in a document already available in the Repository, attaching it is faster and more accurate than asking the Advisor to search for it.

Typing @ in the message field, or selecting Add context from global search, opens the platform search. Files picked there are attached to the message; other results — a vote, a print, a profile — are added as context tags. Several items can be attached at once.

The Advisor receives the identity of each attached item, so it can read the file or record directly and skips the step of searching across all files.

When not to attach a file: when the goal is to find the correct document. A query like "Which of our briefings discusses changes in public procurement law?" should be submitted without an attachment, so that the Advisor searches the files itself.

Reading citations, not just response content

Every data source the Advisor uses is exposed as a verifiable citation chip. Selecting it allows verification of:

• Vote counts and results per club.
• Stakeholder identities.
• The content of a cited post.
• The source document at the cited passage — files open in the file viewer with the Cited passage shown, and legal texts open in the reader at that passage.

The shield below an answer summarizes the source check; its details list the points that need caution. Generated responses generally align with facts, but if a response appears incorrect, run it again with Regenerate response below the answer. The second pass often returns correctly grounded citations.

Reading research maps

A map states what one tile represents and how many results have been assessed. When a map represents parliamentary statements, a completed statement can show its result while other statements are still being collected. Grey tiles have no assessment yet; they do not mean that a statement is unrelated to the topic.

For a yes/no criterion, the percentage is the model's assessment that the criterion applies. It is not the percentage of the statement devoted to the topic. A result of 1% is a low assessed score, not progress. If the method uses the strongest matching passage, other parts of the same statement may discuss different topics.

Select a tile to read its Assessment criterion and assessed text. Evidence opens the original passages and their individual scores. For a statement with several passages, the highest-scored passage appears first when the method uses the strongest match. Check this text before relying on a high score. A research reference chip in an answer or document opens the corresponding map; unavailable results are labelled explicitly.

Utilizing conversation continuity

The Advisor retains the full conversation as context for follow-up questions. Three effective patterns:

• In-progress refinement — "Now break down that same comparison by voivodeship instead of club." — the Advisor reuses the previous result.
• Deepening a specific finding — "Tell me more about the discrepancy in amendment 14."
• Transitioning to action — "Prepare an alert covering the bills you just briefed." — the Advisor takes the identifiers from the previous response and prepares an inactive alert draft for review in the Tracker.

Avoid jumping between unrelated threads in a single conversation. For different topics, open new sessions — the conversation history makes this efficient, and earlier conversations remain searchable.

Steering a running task

A running task does not need to be stopped to change its direction. While the Advisor works, text sent from the message field reaches the running task as a note — for example "Limit it to the current term" or "Leave out the Senate votes". The status below the note shows whether the Advisor has taken it into account.

Use the stop button (Stop research) only when the direction is wrong altogether. What the Advisor has found so far stays in the conversation and can be used in a new, more precise question. See Following the Advisor's work.

Handling partial answers

When a question is too broad to cover in one turn — usually a multi-source question with a very wide scope — the answer covers what it can and ends with a Partial answer. notice that names the gap and suggests how to narrow the question.

This is a valid result. The recommended action is to accept the partial answer and submit a narrower follow-up query, for example limited to one sitting, document, person or period. If partial answers occur regularly for a given question type, the question is likely too broad — two narrower questions yield a better combined result than one very broad one.

Answering clarifying questions

When a tool needs a decision (e.g., "Which Sejm term?", "Which of the two MPs named Kowalski?"), the Advisor asks in the conversation and offers options. Answer with those options and Continue, or Skip the question. Text typed in the message field while the question is open is sent as a note to the running task and does not answer the question.

When the Advisor is the wrong tool

The Advisor deflects several classes of questions:

• General conversation — questions about opinions and evaluations outside the platform's data.
• Out-of-scope analysis — international relations, market forecasts, election polls. The Advisor notes the limitation and answers what is available from the corpus.
• Finished content — the Advisor summarizes, briefs and compares. On request it writes an article draft into the Editor, when the Editor is enabled, but it never publishes or sends anything, and its drafts require review.
• Unrequested or bulk actions — the Advisor researches; its actions are limited and explicit. It prepares drafts for review (alerts, Smart Folders, Editor articles) and changes or deletes an existing alert only on an explicit request, asking for confirmation before a deletion. It does not upload files.

In these cases, the Advisor explains the limitation and points to the appropriate part of the platform when possible.