Advisor skills

The AI advisor's capabilities are partitioned into six skill families, each of which exposes a small set of tools. Every tool call appears in the message timeline as a skill chip identifying which family it came from and what it did.

This page is a reference for what each family is for, the typical phrasings that route to it, and the resource URIs each can return as citations. It is not an exhaustive enumeration of every tool — that surface evolves and is best inspected via the chips themselves.

legislation

The structured record of Sejm activity: votings, motions, sponsors, divisions, and the metadata around them.

Typical questions:

• "What was the result of the most recent vote on the energy law?"
• "How did the budget committee divide on amendment 17?"
• "List the votes from the last Sejm sitting that passed by fewer than five votes."

Returns resources of the form:

• legislation://voting/{proceeding}/{voting_number} — a single vote, opening in Voting ledger.

When the advisor cites a legislation://voting/... URI, the chip in the prose links directly to the seat-by-seat chamber map for that division.

stakeholders

The MP directory: biography, club, constituency, voting attendance, social handles, identified speech segments.

Typical questions:

• "Who is the chair of the Public Finance Committee, and what is their voting record on tax legislation?"
• "Find every MP from the Wielkopolska voivodeship who voted against the climate bill."
• "What did MP Nowak say in plenary about the housing reform?"

Returns resources of the form:

• feature://stakeholder/{stakeholder_id} — opens the stakeholder profile.

The skill family overlaps with legislation — questions about how an MP voted route through legislation, while biographical, contact, and analytical-section questions route through stakeholders. The advisor combines results from both into a single answer when needed.

public_pulse

The aggregated X (Twitter) feed across every tracked stakeholder, with the same filters as the Public Pulse view in the platform.

Typical questions:

• "What has the opposition said publicly about the latest amendment to the labour code?"
• "Has any minister tweeted about the budget vote since it passed?"
• "Surface the trending hashtags from MPs in the past 48 hours."

Returns resources of the form:

• feature://public_pulse/tweet/{tweet_id} — opens the original post on X.

This family is the only one that touches off-platform data. The advisor uses it to ground claims about public positioning — what political actors are saying in public — as opposed to what they have voted for or written in committee.

alerts

The user's existing standing watches and the creation of new ones.

Typical questions:

• "Which alerts have triggered most often this week?"
• "Set up an alert for any mention of the new pharmaceuticals bill."
• "Pause my alert about Brexit-related references — that's no longer relevant."

Returns resources of the form:

• feature://alert/{alert_id} — opens the alert detail.

When the advisor creates an alert in response to a request, it picks tight, topic-specific keywords rather than generic terms. A request for "alert me on energy policy" is interpreted as a starting point — the advisor proposes specific keywords and reports them back so the user can correct the choice if too narrow or too broad.

Alerts created by the advisor are forward-looking: they fire on documents and transcripts ingested after creation, not retroactively.

documents

Search and metadata operations over the user's uploaded Repository files plus the Persate public corpus (politician profiles, party programmes, public documents the platform manages centrally). This family covers both content search (hybrid, filename, sequence) and file-level metadata (enumeration, counts, summaries, metadata reads, full-file reads).

Typical questions:

• "What does our internal briefing say about the gas-pricing mechanism?"
• "Search every government white paper for references to industrial subsidies."
• "Find the passage where the procurement law defines 'critical infrastructure'."
• "How many documents have we uploaded since the start of the year?"
• "Give me a one-paragraph summary of the file we discussed last week."

Returns resources of the form:

• documents://file/{file_id} — opens the file.
• documents://chunk/{file_id}/{chunk_id} — opens the file scrolled to a specific passage.

The advisor selects between three search scopes:

• Tenant + public (default for most questions) — the user's files plus Persate's public corpus.
• Public only — when the question is about general legislative material, not the user's own documents.
• Tenant only — when the user's question is explicitly about their own files ("in our internal briefings, ...").

Where a file is mentioned by name, the advisor falls back to filename or sequence search rather than the heavier hybrid pipeline.

chat_context

Lookup against the user's prior conversations with the advisor.

Typical questions:

• "Earlier today we were discussing the energy bill — pick that thread back up."
• "What did you tell me last week about the procurement reform?"

Returns resources of the form:

• chatcontext://session/{session_id} — opens a past conversation.
• chatcontext://message/{message_id} — opens a past message in context.

This family enables continuity across sessions without the user having to manually re-paste prior context. It is read-only from the advisor's perspective: it cannot delete, edit, or annotate prior conversations.

How the advisor decides which to call

The model selects skills based on the question's surface form and the manifest of available tools. There is no explicit routing the user controls — the routing is the model's job. In practice:

• A question that names a specific MP almost always pulls stakeholders.
• A question about what someone said publicly pulls public_pulse.
• A question about how someone voted pulls legislation.
• A question about a document — whether by filename, topic, or content — pulls documents.
• Multi-part questions pull from multiple families and braid the results.

When the advisor encounters a question that has no matching tool — for example, a request for sentiment analysis or for material outside the legislative corpus — it answers what it can answer and notes the unanswered part rather than aborting the whole turn.