--- title: "How AI uses and cites sources" description: "Text and semantic search, answer generation, access controls and the limits of citations." language: en canonical: https://cercaposta.it/en/docs/explanation/ai-sources/ updated: 2026-09-19 publisher: WpWeb S.r.l. diataxis: explanation --- # How AI uses and cites sources CercaPosta AI chat searches the archive to retrieve information and build answers linked to sources. The model does not need to have memorised your emails during training. Relevant documents are supplied in the request context, within applicable permissions and AI preferences. ## Retrieval and generation are different steps Classic search finds terms and applies filters. Semantic search uses numerical text representations, called embeddings, to recognise similarity in meaning. CercaPosta can combine these approaches and rank relevant passages before generating an answer. RAG, or retrieval-augmented generation, connects an answer to information in the archive. Results depend on both retrieval and the model: incomplete retrieval can produce an incomplete answer even when the right document exists. ## What a citation establishes A citation lets you open the referenced message or attachment. It helps check the passage, date, sender and context. By itself, it does not establish that every sentence in the answer is correct or that all possible sources were considered. For contracts, deadlines or decisions, check that the cited text supports the specific claim. A draft and a final version may coexist; a statement may have been superseded by a later reply. ## Data scope The plan, personal preferences, source or folder rules and access permissions limit what a feature can use. Configuring a model does not give it unrestricted access to every archive. The installation’s selected AI endpoints also determine where context is processed. See [choose which emails AI can use](https://cercaposta.it/en/docs/how-to/control-ai/) and the [privacy policy](https://cercaposta.it/privacy?lang=en) for settings and data handling. ## Public documentation and private archives This documentation describes the product and is publicly readable, including in Markdown. Allowing crawlers to read the website concerns these public contents: it does not authorise access to users’ mailboxes or archives. Training a model on public content and using an archive to answer a question are separate processes. A public page does not guarantee that an external assistant knows the product or describes its features correctly. Current documentation and service conditions remain the sources to consult. ## Practise source checking Follow the [first AI answer with sources](https://cercaposta.it/en/docs/tutorials/first-ai-answer/) tutorial: use a familiar fact to practise checking the original reference. ## An example verified in the app In the `FA-2026-0001` agreement exercise, the AI answer stated a 60-day notice period and a 12-month renewal. We opened citation [1] and the attached agreement, which contains both terms. The screenshot shows that original PDF; all names, messages and documents in the demo are fictional. In the Italian version of the exercise, we opened the confirmation email referenced by citation [4]. Different sources can support the same answer: check the passage actually cited instead of expecting a fixed citation number.

**Sample data · real screenshot**The original agreement PDF opened from the AI answer’s citation. · Open the full-size screenshot