# How Ask Siri constructs and uses context > Source: > Published: 2026-10-17 07:00:00+00:00 For all AI, the prompt and additional instructions provided are critical, and one of Siri’s great challenges. If Ask Siri is to be usable, it has to resolve the ambiguities in your queries without requiring you to provide lengthy explanations, or engaging you in a long series of clarifications. This article illustrates how it does that in macOS 27.0.1 Golden Gate. English is a language built on ambiguity, with simple words like *really* that can mean anything from certainly to the exact opposite, and *no* that can mean yes. We often need the additional clues given by intonation in speech, although we can only see the words in writing. For almost every word, context is critical to understanding. Siri’s ability to use context to eliminate ambiguity is neatly illustrated by asking it a basic, and thoroughly ambiguous, question, such as “How far is it by road from Portsmouth to London?” If you don’t know where I am, there are too many possibilities, with London in England, Ontario, Ohio and others, and Portsmouth in England, New Hampshire and Virginia. But knowing that I’m only a few miles from Portsmouth in England, and the only London I could drive to from there is the country’s capital city, Siri makes the right call in that context, without having to quiz me on geography. Siri’s answer exceeds expectations, even anticipating some of my likely follow-up questions. The best way to observe how Siri constructs and uses context is by analysing an Apple Intelligence Report, or AIR, as you Ask Siri some questions, such as those I used [earlier this week](https://eclecticlight.co/2026/10/14/where-to-look-for-mac-help-tips-or-ai/). The prompt and instructions actually passed to the model running on Private Cloud Compute (PCC) are thorough and long-winded, a total of over 16,000 words before it even gets to my query. In the quotations below, I have tweaked some of the punctuation to make it more readable, and omitted some of the system overhead for clarity. Analysis was performed using my free [AIRer](https://eclecticlight.co/consolation-t2m2-and-log-utilities/). It opens by establishing the AI’s role and persona: “You are Siri, an intelligent assistant designed by Apple in California. You lead with the answer in `[standalone_answer]` — the information the user asked for comes first, with optional supplementary depth after via `[supplemental_material/]`. Your role is to understand, handle and respond to requests from the user using available tools and resources.” There are explicit instructions covering areas that could prove ambiguous: “Missing or insufficient info: issue an `ask_user` call (asking in plain response text is not equivalent) whenever progress needs information only the user can provide — a missing parameter, an ambiguous reference, or a choice between paths. Omit optional parameters the user didn’t provide.” “Ambiguous targets: use `ask_user_to_pick` before proceeding, with enough detail to distinguish. Resolve silently only when context is decisive — one result exists, the conversation singles one out, the user said “that one” or referenced something just discussed, or recency eliminates alternatives. Always ask when multiple entities stay plausible, names only partially match, several contacts share a first name, the action is irreversible, or nothing breaks the tie. When in doubt, ask.” Those final four words are crystal clear. One potentially valuable source of context is the text currently displayed in the active window. In this case, it was passed as `"kind":"OnScreenText", "is_editable":true, "is_focused":true, "missing_details":true,` "text":"How can I keep files downloaded in iCloud Drive? How can I copy the login keychain from one Mac to another? How can I roll back to a snapshot of my Data volume?" This emphasises the importance of the privacy provided by PCC. That’s not the sort of information you’d want to hand over to Google, for example. More general context was also passed: `{"current_time":{"kind":"DateTimeEntity", "level_of_detail":"full", "iso_timestamp":"2026-10-13T10:32:53+0100", "weekday":"Tuesday"}, "current_user":"E Howard N", "date_and_time":"24-Hour Time", "device_type":"Mac", "focused_app":{"kind":"Application", "level_of_detail":"full", "foreground_window":{"kind":"UIWindow", "level_of_detail":"full", "entities"[{"capped":false, "total_count":"1", "count_shown":1, "level_of_detail":"full", "entities_level_of_detail":"minimal", "entities_level_of_details_available":["full"], "entities":[{"kind":"OnScreenText", "is_editable":true, "is_focused":true, "missing_details":true, "text":"How can I keep files downloaded in iCloud Drive?` How can I copy the login keychain from one Mac to another? How can I roll back to a snapshot of my Data volume?"}], "entities_kind":"OnScreenText", "missing_details":true}], "screenshot" {"kind":"Visual", "level_of_detail":"full", "level_of_details_available":["full"]}, "title":"DelightEd"}}, "language":"English", "locale":"en-GB", "region":"United Kingdom", "response_mode":"Display+Voice", "span_matches":{}, "user_gender":"unspecified"} Then the query is passed. Note that this is the same text as my original question, but Siri has added the two words *on Mac* to sharpen the context: `"How can I roll back to a snapshot of my Data volume on Mac"` Prompted with that mass of text, the model then gets to work to respond to that query. One of the first things it does is perform a find, including a web search: `response:default_api:find{results:[{"results":[{"global_entities":[{"kind":"WebSearchResults",` That returns ten results, starting with: `"num_results_shown":10, "num_total_results":15, "query":"How can I roll back to a snapshot of my Data volume on Mac", "results":[{"kind":"WebPassageEntity", "body":"Time Machine lets you restore files from local snapshots of the files on your Mac, even when your Time Machine backup disc isn't available.` Results are then scored for relevance, and include a link to the article on this blog: `"score":"0.72", "title":"How to make and roll back to a snapshot", "url":"https://eclecticlight.co/2026/05/07/how-to-make-and-roll-back-to-a-snapshot/"}, {"kind":"WebPassageEntity", "body":"Snapshots are a feature of APFS ([link id=\"entity_link_Q24577174\"]Apple File System[/link]) that has been part of the [link id=\"entity_link_Q14116\"]macOS[/link] since High Sierra....` Although at least some of those are discovered from external sources, it’s also possible that some are derived from local browsing history. I’ll return to that another time. From those, Siri composes its response, and it’s passed back for display. Although I had expected the general context to be provided, the text content extracted from the displayed window came as a surprise. It was consistently passed to the model in all such Ask Siri exchanges I have examined so far. No wonder Siri appears to contextualise so effectively with such brief queries.