Fetch the news of a topic in an inclusive date range from every configured source.

The five feeds run in parallel and the result is merged, de-duplicated and filtered to the range. A source that fails is skipped with a warning, so a broken feed thins the result instead of failing the call; if no source is configured at all (no GDELT, which needs no key) the call is a tool error.

Read the numbers, dates and symbols out of a text with the extraction agent.

The model's reading is combined with the deterministic regex pass, so a mention the model formats differently ("eleven" beside "11") is still found, and a model that cannot answer at all degrades to the regex reading instead of failing. sources says which strategies contributed. Needs an LLM endpoint (OPENAI_API_KEY or OPENAI_BASE_URL).

Reduce a text to its numerological value.

Returns the raw gematria sum, the digit-reduced value (11, 22 and 33 are master numbers and are not reduced further), whether the value is master, and the rendered calculation steps. Pure logic: no Qdrant, no model, no configuration.

Find the connections among the stored news items with these ids, save them and return them.

The ids come from an earlier fetch_news/ingest run or from query_qdrant. An id that is not in the store is skipped rather than refused, and an empty list answers [] without asking the model — "nothing connects" and "nothing to look at" stay tellable apart. Each returned pattern carries the discovered_at timestamp save_pattern stamped. Needs Qdrant and an LLM endpoint.

Report the master numbers (11, 22, 33) among a list of numbers.

master_numbers lists the distinct master numbers found, ascending, and count how many occurrences there were: [11, 11] gives has_master=True, master_numbers=(11,) and count=2. An empty list is a valid input and reports no master number. Pure logic: no configuration, no service.

Return the numerological reading for a calendar day, saving it for later calls.

A day that was already read answers from Qdrant without running a model. Otherwise the reading is assembled from the last seven days of stored news (their dominant number, whether a master number is active, and the patterns among them) plus the activations of exactly those numbers over the pipeline's thirty-day memory window, and then written back. Needs Qdrant and an LLM endpoint.

Search the stored news or patterns by meaning and return the matching entities.

collection="news" runs the hybrid (multi-stage) search over stored articles and honours the optional payload filters — publication window, source, reduced value or master number. collection="patterns" finds saved pattern interpretations and takes no filters: passing any is an error rather than a silently ignored argument. Needs Qdrant only; nothing is embedded remotely and no model is called.

The returned items are the typed entities themselves (NewsItem or Pattern), best match first; the scores stay in Qdrant because the callers work with the entities.

Store one pattern and return it as written, with discovered_at filled in.

The original design sketch called the return type SavedPattern; the saved model is a Pattern (the one numenews.vector.save_pattern returns), so the tool returns that rather than a second type saying the same thing. Re-saving the same pattern id overwrites the point instead of adding a copy, and a timestamp already present is never rewritten. Needs Qdrant only.

Return when a number was activated in the news, newest first, inside the last days days.

One entry per (news item, number) pair the ingest stored: the day the article was published, the item's id and the snippet the number was read in. The window ends today and includes it (days=1 is today). The original design sketch wrote the signature as get_history(number); the read is a window, so days is a parameter whose default of 30 matches the memory window. Needs Qdrant only.