NEWS

textpress 1.1.2

fetch_urls() and its DuckDuckGo scraping implementation have been removed.
fetch_rss() adds publisher-supported RSS and Atom discovery. It returns a flat table of article URLs and feed-supplied metadata that can be filtered before the selected URLs are passed to read_urls().
fetch_rss() assigns stable doc_id values. Passing its result table directly to read_urls() preserves discovery metadata in the returned meta table.
rss_politics adds a verified collection of 67 active, non-government US politics feeds from 61 sources. The exported catalog and its canonical CSV cover general political news, Congress, elections, polling, public opinion, and state politics; it is available directly as textpress::rss_politics.
rss_local_rags adds a pinned Local Rags RSS snapshot with the shared feed fields plus Census division, state, county, five-character FIPS geography, 2023 Rural-Urban Continuum Codes, and metro status. The local catalog omits the politics-specific category and source_type fields. Detailed discovery and 3DLNews provenance remain in the source repository.
fetch_wiki_refs() now always returns one flat table for single or multiple Wikipedia URLs, allowing its results to pass directly to read_urls().
Wikipedia requests now identify textpress with a package User-Agent.

Six vignettes added covering the full pipeline: web data, Wikipedia data, regex search, dictionary search, semantic search (RAG), and basic NLP processing.
Basic NLP vignette walks through nlp_split_sentences(), nlp_tokenize_text() (word and Biber methods), and nlp_cast_tokens() stepwise and as a single pipe.
README revamped: tighter intro, API map, RAG/agent positioning, vignette links.

util_fetch_embeddings() re-added for embedding generation via Hugging Face inference endpoints (reversed 1.1.0 removal; now calls the HF inference API rather than loading models locally).
nlp_cast_tokens() documented and surfaced -- flattens the token list from nlp_tokenize_text() into a long-format data frame with optional character spans.
Suggests trimmed: ellmer and unused packages removed.

Package is now organized around a four-stage pipeline: Fetch → Read → Process → Search. All functions use a consistent verb_noun pattern.
Acquire: fetch_urls() (from web search), fetch_wiki_urls(), fetch_wiki_refs() — return URLs or metadata, not full text.
Ingest: read_urls() — read content from URLs into R (replaces web_scrape_urls).
Process: nlp_split_*, nlp_tokenize_text(), nlp_index_tokens() (and nlp_roll_chunks() for rolling windows).
Search: Four retrieval options — search_regex() (regex/KWIC), search_index() (BM25), search_vector() (cosine over your own embeddings), search_dict() (dictionary match; replaces ner_extract_entities).
Common parameters standardized: corpus (replaces tif), by (replaces text_hierarchy).

In-package embedding generation (e.g. Hugging Face API). Use your own embedding pipeline and pass your embedding matrix as the \code{embeddings} argument to \code{search_vector()}.
Legacy names: web_search, wiki_search, wiki_find_references, web_scrape_urls, ner_extract_entities, sem_nearest_neighbors / sem_search_corpus (replaced by search_vector and search_regex).

README revamped around the API map and a single “golden path” workflow.
DESCRIPTION and package help updated for the four-stage pipeline; version set to 1.1.0.

Initial release: URL fetching, URL content reading, NLP processing (split, tokenize, index), and corpus/search utilities.