Filtering NVIDIA news on the headline alone drops 3,890 of 5,718 articles in a week

The finished watcher runs two queries per window because neither filter contains the other. One asks the API for articles whose title holds the string, `title=nvidia` on the `/v1/news/everything` endpoint ^[19]. The other asks for the canonical entity id, which matches regardless of spelling and carries NVIDIA, nVidia and NVDA as aliases in the entity metadata ^[11][12]. Named-entity recognition misses articles too, according to the post, especially on smaller and non-English outlets ^[14]. Two requests a window is the cheapest line in this design ^[10].

Stage three is where the volume goes. Only 32.2% of entity-matched articles put the company in the headline; the rest name it in passing, in a monitor review or a fund's holdings list ^[9]. The rule is the headline plus at least two mentions in the body, counted off `entities[].frequency` ^[11], and it took 258 articles a day down to 97, a cut of 62.4% ^[9][5].

Then the same story arrives again. 12.5% of what survives filtering is a republication under a near-identical headline ^[8]. At 97 a day, that is about 12 ^[6]. The response already carries `is_duplicate` and a `story.id`; in the sample record printed with the post, the story id equals the article id and `is_duplicate` is false ^[12].

1,806 of the 5,718 articles were in English, 31.6% ^[6], so 3,912 were not ^[2]. German was 9.4% and Spanish 7.4% ^[7]. Leaving `language=en` unset leaves English theme keywords matching against text they cannot match ^[16]. "I set it, because I can't read the rest. That's a choice, not a default," the author wrote ^[13].

The recall figures do not close. The post says a headline search finds 29.8% of the coverage ^[3], which is 1,704 of 5,718; it also says the headline query misses 3,890, which leaves 1,828 ^[3]. The 180 headline hits with no entity extracted are given as 9.8% of the headline set, implying about 1,837 ^[5][4]. Low figure to high figure, that is a gap of roughly 124 articles ^[3].

For the counts to transfer, your subject has to look like this one. NVIDIA's entity record is flagged `is_public_entity` and carries a ticker alias ^[12], and the week's coverage ran in at least eight languages ^[7]. A private domestic supplier gets a different language mix and, I would expect, worse extraction than a 9.8% miss. That 9.8% is one provider's NER over one week, measured 2026-09-06 ^[15][5]. The one sample record the post prints is an earnings story dated 2026-09-02 ^[17], so 817 a day is not a quiet-week baseline.

Version one of the script was fifteen lines, and the author wrote that it was wrong in a way the output did not show ^[18]. Five stages later, 161 of the 5,718 articles were left, 2.8% of the pull ^[2][1].