--- title: "AI Processing" id: "1958" type: "page" slug: "ai" published_at: "2026-09-30T22:50:44+00:00" modified_at: "2026-09-30T23:50:35+00:00" url: "https://xedant.com/agents/news/docs/ai" markdown_url: "https://xedant.com/agents/news/docs/ai.md" excerpt: "Collection yields raw material: pages, paragraphs, numbers. A human cannot read that, so the AI…" --- # AI Processing [https://xedant.com/agents/news/docs/ai.md](https://xedant.com/agents/news/docs/ai.md) Collection yields raw material: pages, paragraphs, numbers. A human cannot read that, so the AI stands between collection and you. Its job is to turn hundreds of items into short, clear facts — and to stay honestly silent where it knows nothing. ## Why AI - **Shorter.** Instead of a page of text — two sentences and a link to the source. - **Clearer.** Instead of “−25.6%” — “the price fell by a quarter”. - **More connected.** Material about one happening gathers into one story instead of arriving in scatter. - **More discerning.** The materiality rating separates what deserves your attention from the background. ## A summary built strictly on the material The main rule of processing: **no inventions**. The summary is built only from the collected material, and next to every statement stands a link to the exact quote. When there is no backing in the text, an honest note that no confirmation was found appears instead of a pretty retelling. This matters in news: an invented phrase sounds exactly as convincing as a real one, and is discovered only when it is already too late. ## The materiality rating The rating runs on a scale from 0 to 10 and then sorts into four clear bands — “noise”, “worth knowing”, “material”, “existential”. The threshold where “material” begins is yours to set; by default the product is tuned not to send you trifles. When a rating looks wrong, tell the agent: it will fix the rules or add a caveat. An unprocessed event is honestly marked “unscored” — there are no invented numbers in the interface. ## Topics Every item automatically gets topics from your own structure: “prices”, “supplies”, “regulation”, “our products”. Topics are then handy for filtering events and building rules: “everything about regulation — into a separate digest”. ## Entities Entities are the companies, people and products mentioned in the material. The agent keeps their list and marks the mentions: you can see how many times a competitor came up in a month, and in what context. The list is edited in words: “add this company to the list, it is our supplier”. ## Stories When one piece of news has been reprinted by ten outlets, reading ten items is pointless. The agent gathers material about one happening into a **story** — a page showing how the news developed. - **New** — the story has just appeared; no follow-ups yet. - **Developing** — new material keeps arriving. - **Fading** — no new material for a long time. - **Archived** — the story was closed by you or by the agent. ## Translation Material in a foreign language can be translated on request: the translation is shown next to the original, so the wording can always be checked. Translation is done on your request, not for everything at once — that keeps AI spending under control. ## Confirmation by independent sources The agent counts how many independent sites have reported one happening. The confirmation count is visible in the story, and a protection against lone news items can be built on it: a rule like “do not send until at least two different sites confirm the news” is described in the [Alerts & Noise Protection](/agents/news/docs/alerts) section. ## Your own AI provider and your own processing Processing can not only be configured but also replaced with your own Python code: - **Any AI task can be routed to your own code** — the summary, the materiality rating, the topics. Useful when things must be computed by your own rules, or without calling an external service. - **Quotas and spending accounting work the same** as with ordinary models: you see how many tasks went out and how long they took. - **You can add your own facts** — the code walks all the items and events and leaves its own marks, on which rules then work: “act when the code counted this specific fact”. - **Unavailable code is not replaced by someone else’s answer** — the task is honestly marked as not completed. Details — in the [Custom Scripts](/agents/news/docs/scripts) section. ## Queue, quotas and spending Every model call is money and time, so processing goes through a queue with two limiters: - **The daily call quota** — how many AI tasks are allowed per day. The quota ends loudly: the interface shows that processing has stopped and why. - **Spending accounting per task** — you see how much went to summaries, how much to materiality ratings, how much to publication drafts. When the quota is exhausted, the collected material and events stay in place — processing simply waits. Nothing is lost, and analytics shows exactly where you hit the limit. ## What the AI does not do - **It does not detect the mood of a text.** The function exists, but it honestly answers “not configured” instead of inventing percentages. - **It does not transcribe videos, podcasts and pictures.** Such material keeps only its descriptions. - **It does not publish on its own.** The AI prepares a draft, and sending happens only after your approval: [Publishing](/agents/news/docs/publishing) . Next → [Rules](/agents/news/docs/rules) [← Back to the documentation index](/agents/news/docs)