How it works
Method and editorial choices
In short
Fascicoli's answers are generated by AI models from the official verbatim records of the Italian Chamber of Deputies and Senate (19th legislature). Every quotation is verified word for word against the source; the selection of sources and the synthesis are choices made by the system, governed by the criteria described on this page.
From question to answer
Each question goes through a staged pipeline: the query is rewritten and expanded, the system retrieves relevant speeches through hybrid semantic search, computes each MP's authority on the topic, generates a section for every parliamentary group and integrates them into a single answer. Quotations are extracted by position from the original text: if a quotation does not match the record, it is discarded.
The “Behind the scenes” panel available on every answer shows the stages, timings and model calls that produced it.
The authority score
An MP's authority on a topic is a weighted sum of six components, each between 0 and 1 and computed against the question asked: there is no global score, only a score per topic.
| Criterion | Weight |
|---|---|
Floor speeches How much, and how recently, they spoke about the topic on the floor | 25% |
Committee Relevance of their committee to the topic | 25% |
Legislative acts Acts signed on the topic, weighted by recency | 20% |
Profession Affinity between professional background and topic | 15% |
Education Affinity between educational background and topic | 10% |
Institutional role Current institutional offices | 5% |
These weights are an editorial choice, not a neutral measure: they declare which signals count and by how much. We publish them so they can be inspected and challenged; on the authority analysis page you can change them and watch the ranking shift.
Exact quotations, declared selection
Every quotation is extracted by position from the verbatim record and verified word for word: what appears in quotation marks was actually said, and each quotation links to the official document.
Exactness does not remove selection: which passages to quote, and in what frame, is the system's decision. An authentic quotation can still be partial with respect to the full speech or the speaker's overall position. The full context is always one click away, in the official record.
Limitations
Answers are generated content and can contain errors or omissions: a relevant speech may not be retrieved, a position may be summarised incompletely. ParliamentRAG is an independent research project, neither affiliated with nor endorsed by the Chamber of Deputies or the Senate. For anything that matters, check the official record.
The ideological compass
For each question the system generates two thematic axes and asks a language model to classify every retrieved speech: a position between −1 and +1 on each axis, or no position at all. A group's position is the mean of only its speeches that take a stance on that axis, attenuated toward the centre when they are few: a sharp position has to be earned with numbers, not with three sentences.
Two reading cautions. A group's coordinates can rest on different sets of speeches for the two axes, and the number of speeches behind each coordinate is shown in the tooltip. Points at the centre take no position: the compass describes what the groups argued in the Chamber on this question, it does not certify their political placement.
Going deeper
The system's architecture, weights and evaluation are documented in the scientific paper: