Best practices
Practical guidance for getting the most out of the AI advisor — choosing a depth, framing questions for the corpus, when to attach files, how to use citations, and the failure modes to recognise.
The AI advisor rewards a particular style of question and penalises others. This page distils the patterns that recur across heavy users into operational guidance.
Choose the depth that matches the question
The depth dial is the single most consequential control. Get it right and the advisor returns appropriately — fast for lookups, thorough for analysis. Get it wrong in either direction and the experience degrades.
| Depth | Choose when... | Avoid when... |
|---|---|---|
| Surface | The answer is a single fact or a small list, the source is obvious, and you'll either accept the first plausible result or rephrase. "What was the result of the latest energy vote?" | The question requires synthesis across multiple sources or a comparison. Surface depth will under-search. |
| Balanced | The everyday setting. The question requires the advisor to ground its answer with a few lookups, and you want a complete answer reasonably quickly. | A Surface answer would suffice (you're paying for needless lookups). |
| Deep | The question is genuinely multi-source — comparison, briefing, opposition research, or "surface anything noteworthy". You're prepared to wait longer for a substantively richer answer. | The question is narrow. Deep depth will over-search and may return more than you wanted. |
A useful heuristic: if the answer fits in one short paragraph, Surface is enough. If it needs section headers, Deep is justified. Balanced covers the everything in between.
Frame the question in terms of the corpus
The advisor is anchored to specific entities — Sejm proceedings, MPs, parliamentary clubs, voting numbers, file names. Questions framed in those terms produce sharper answers than questions framed in abstract terms.
- Less effective: "What's happening with energy policy?" — the model has to infer scope.
- More effective: "Summarise the most recent floor activity on the renewable-energy bill, including the latest vote and any committee actions in the past week." — the model knows what to search for.
This does not mean every question must be ten lines long. It means anchor at least one concrete entity — a bill name, an MP, a club, a date range, a committee, a file. The advisor's tools are organised around those entities, and naming one of them routes the question to the right skill family immediately.
Attach a file rather than describing it
When the answer is in a document the user already has, attaching it is faster, cheaper, and more accurate than asking the advisor to find it.
- Press
@in the input to open the file picker. - Search by filename if you know the name; toggle to hybrid search if you only remember what the document was about.
- Attach as many files as relevant; the advisor handles multi-file context.
The advisor sees an attached file as a first-class reference — it bypasses the search step and goes straight to extracting from the file. This matters more than it sounds: a Hybrid search across the entire corpus is the single most expensive call the advisor makes.
When not to attach: when you want the advisor to find the right file. "Which of our briefings discusses the procurement-law changes?" should be asked without an attachment so the advisor's hybrid search runs.
Read the citations, don't just trust the prose
Every URI the advisor surfaces is auditable. The chips are not decorative.
- Click
legislation://voting/...chips to verify the vote count and the per-club division. - Click
feature://stakeholder/...chips to verify the MP is who the advisor named. - Click
feature://public_pulse/tweet/...chips to confirm the post says what the advisor paraphrased. - Click file citation superscripts to download the source file at the cited passage.
Where a chip seems off — the wrong MP, the wrong vote, the wrong file — regenerate the response. The advisor's stochastic nature means a second pass often produces correctly grounded citations even on a question it just got wrong.
Use conversation continuity
The advisor retains the full conversation as context for follow-up questions. Three patterns work well:
- Refine inline. "Now break the same comparison down by voivodeship instead of by club." — the advisor reuses the prior result and only re-runs what changed.
- Drill into a specific finding. "Tell me more about the divergence on amendment 14." — the advisor goes back to the relevant tools.
- Pivot to action. "Set up an alert covering the bills you just briefed me on." — the advisor pulls the bill identifiers from its prior answer and creates the alert without re-asking.
Avoid thread-switching — a single conversation that jumps from energy policy to procurement reform to a personnel question loses focus. Open a new conversation for unrelated topics; the chat history makes it cheap.
Recognise the budget-exhaustion signal
When a question genuinely exceeds the per-turn budget — typically a Deep query with too broad a scope — the response ends with a brief note about what could not be completed.
"...Of the eleven bills you asked about, I covered nine in detail; for bills 10 and 11 I ran out of turn budget. Re-run the question scoped to those two specifically and I'll cover them in the next turn."
This is the correct outcome — better than a hard failure mid-stream. The remediation is to:
- Take the partial answer at face value (it is grounded).
- Reissue the unfinished part as a follow-up question, narrower.
If you find yourself hitting budget exhaustion regularly on the same kind of question, the question is probably too broad. Splitting the question into two or three Balanced-depth turns will produce a better total result than a single Deep turn.
Don't replicate elicitation in your prose
When a tool needs disambiguation, the advisor surfaces an inline form ("Which Sejm term?", "Which of these two MPs named Kowalski?"). Answer the form rather than re-prompting in the chat.
- Less effective: seeing the disambiguation form, ignoring it, and typing "I meant the current Sejm, term 10" into the chat input. The form remains unresolved; the advisor sees a new turn rather than a continuation.
- More effective: answering the form. The original tool resumes and the response continues from where it paused.
When the advisor is the wrong tool
The advisor is not built for, and will redirect from, several classes of question:
- General-purpose chat. "What do you think of the Polish electoral system?" The advisor is constrained to factual, sourced answers about specific Sejm activity; it will not opine.
- Off-corpus political analysis. Questions that require context from outside the legislative corpus — international relations, market forecasting, electoral polling — are not what the toolkit covers. The advisor will note the limitation and answer what it can from the corpus rather than inventing material.
- Drafting. The advisor will summarise, brief, and compare, but it will not draft a press release, a legal brief, or a position paper. Use the summarised material as input to your own drafting.
- Bulk operations. "Send an email to every MP in the Energy Committee." The advisor performs queries, not actions. Where actions exist on the platform — alert creation, file uploads — they are explicit and confirmed; there is no mass-action surface.
For these cases, the advisor will produce a brief reply explaining the limitation and (where possible) point to the part of the platform that fits.