AI advisor
Persate's in-app legislative-intelligence assistant: an agent that searches the Sejm corpus, queries the platform's structured data, and writes grounded, cited answers in plain language.
The AI advisor is Persate's in-app assistant for exploratory work over the legislative corpus. It is not a free-form chatbot — it is an agent with a fixed-purpose toolkit, scoped to the Polish parliament and the documents in the user's account. Every claim it makes is anchored to a real lookup, and every reference is a clickable citation back to the source.
It is accessed via the sidebar entry INTELLIGENCE → Advisor.
What the advisor is for
The advisor fits between two tasks the user would otherwise do manually:
- Targeted retrieval that doesn't quite fit a structured search. "Which committee considered the latest energy-pricing bill, and who voted against?" would otherwise require navigating the Voting ledger by hand and cross-referencing committee transcripts. The advisor performs the cross-reference and cites both sources in a single response.
- Synthesis across sources. "Compare what the Climate Coalition has said publicly on coal phase-out with how its members actually voted in the last six months." draws on Public Pulse for the public statements and the Voting ledger for the divisions. The advisor pulls both, aligns them, and presents the comparison.
Tasks the advisor is not built for: open-ended drafting, code generation, mathematical reasoning detached from the corpus, or general-purpose conversation.
How it differs from a generic chat
Three differences shape the experience.
- Grounded answers. The advisor cannot answer questions about the Sejm from memory. To produce any factual claim it performs a lookup against the platform's data — voting records, MP profiles, recorded speeches, uploaded documents — and the chips visible above each answer are those lookups.
- Multi-source citations as inline chips. Every reference the advisor makes appears in the response as a clickable badge that opens the corresponding record in the platform. Citations are not optional; the advisor is constrained to cite at most one source per atomic claim.
- Bounded turns. Each user message produces one turn, and every turn is bounded so it cannot run away. When the advisor approaches the limits of what it can do in one turn, it wraps up gracefully — the answer arrives, sometimes flagged as incomplete on the difficult sub-questions, rather than spinning indefinitely.
What the advisor can do
The advisor's capabilities are partitioned into six skill families, exposed via the platform's MCP layer:
| Family | What it covers |
|---|---|
| legislation | Sejm voting records, motion metadata, division outcomes; per-MP voting history. |
| stakeholders | The directory of MPs — biography, club, contact details, social handles, attendance stats. |
| public_pulse | Public-pulse activity (X/Twitter) from tracked stakeholders. |
| alerts | The user's existing alerts, plus creating new ones. |
| documents | Hybrid, filename, and sequence search over uploaded files and the platform's public corpus, plus per-file metadata, listings, counts, and summaries. |
| chat_context | The user's prior conversations with the advisor, used to maintain context. |
These areas are described in more detail in Advisor capabilities.
Thinking depth
A small dial at the left of the input controls the thinking depth applied to the next message:
- Surface — fast and direct; favoured for follow-up questions and lookups with an obvious single answer.
- Balanced (default) — the everyday setting; the advisor grounds its answer and follows up on obvious gaps without exhausting every lead.
- Deep — thorough; the advisor explores related material, cross-references, and pursues lateral leads. Slower per message.
The depth is set per-message; the default for new messages is configured in Settings → Advisor. Changing the per-message dial does not change the default. The full mechanics are described in Conversations and Best practices.
Conversation history
The right column of the screen lists every conversation on the user's account, most recent first. Each entry is identified by the topic the platform inferred from its first messages.
- New conversation clears the workspace and begins a fresh session.
- Selecting a past conversation reopens it — the full transcript is reconstructed including the original lookups and citations.
- The trash control on each entry deletes that conversation and its transcript permanently.
History is per-user.
Refresh-and-resume
A turn that is in progress when the user navigates away or refreshes the browser is not lost. Reopening the conversation reattaches to the in-flight response and replays whatever has happened so far before it resumes streaming.
In practice this means it is safe to send a deep query, switch to Repository to attend to something else, and return — the answer will be waiting (or still streaming) when the conversation is reopened.
Pages in this section
- Conversations — composing messages, mentioning files, the inline depth selector, the message timeline, citations, copy and regenerate, and history.
- Advisor capabilities — what the advisor can do, with examples of the kinds of question that route to each area.
- Worked examples — substantive prompts and the expected behaviour, drawn from the recurring use cases of public-affairs and legal teams.
- Best practices — choosing a depth, when to attach files versus letting the advisor search, citation handling, and the pitfalls of generic queries.