The library SeferAI searches just got twenty-four times bigger — and it reads Hebrew

Until this week, SeferAI searched 156,185 passages of Jewish text. It now searches 3,802,528.

That is not a marketing number, and it is worth explaining what actually changed, because the interesting part is not the size.

Hebrew was never really in there

The old index held English only, and — more to the point — the model that matched your question to a passage had been trained on English alone. Ask it a question in Hebrew and it did something that looked like searching and wasn't.

Here is a real example. The question was הלכות נר שבת — the laws of the Shabbat candle.

Before, the top result was Bartenura's commentary on Mishnah Maasrot, which is about agricultural tithes. The next four were no better.

Now, the top results are Kiryat Sefer on Mishneh Torah, Shabbat, and four passages of Peninei Halakhah on Shabbat.

The new index holds 3,118,549 Hebrew passages beside 683,979 English ones, and the matching is done by a multilingual model that reads both. A question in either language can now reach a source in either language.

Where the depth actually went

Almost all of it is commentary — 2,365,241 passages across 6,116 texts. Asking what the Rishonim make of a verse now has six thousand commentaries behind it rather than a sample of the ones that happened to be translated.

| | texts | passages | |---|---|---| | Commentary — Rishonim, Acharonim, modern | 6,116 | 2,365,241 | | Halakhah | 383 | 227,592 | | Talmud | 187 | 204,158 | | Liturgy | 87 | 196,662 | | Chasidut | 162 | 170,009 | | Responsa | 163 | 140,279 | | Jewish Thought, Kabbalah, Midrash, Musar | 449 | 349,912 | | Tanakh, Targum, Mishnah, Tosefta, Second Temple | 528 | 148,675 |

Tanakh barely moves. It was always mostly there. What changed is everything written around it.

A new filter: Liturgy

Prayer books are now searchable — siddurim, machzorim, selichot — but as a checkbox you tick rather than something that is always on.

The reason is worth being straight about. When we first put the liturgy in with everything else, a question about mystical union with the divine came back as four nearly identical passages of Selichot, because the same prayer repeats across different nusachim. A siddur is the text of a prayer; it is usually not the source that answers a question about the prayer. With liturgy set aside, that same question returns Ma'amar Zohar HaRakia, Ohr Ne'erav and the Tanya — which is the answer.

So it is there when you want it, and out of the way when you don't.

One thing that got worse — and the dial you have

More sources is not automatically better answers, and one of our test questions got worse. What does the Torah say about honoring parents? used to return Exodus 20:12 and Kiddushin 31b — the commandment itself. It now returns thoughtful secondary literature instead.

The verses have not gone anywhere: Exodus 20:12, Deuteronomy 5:16 and Kiddushin 31b are all in the index, in Hebrew and English. It is a ranking problem, not a coverage problem. The new model prefers a paragraph that discusses an idea to a verse that states it, and against three and a half million passages a short verse has a great deal of company.

Which is what the source checkboxes are for, and they are worth more now than they were last week. Ticking a narrower set does not just filter the results — the filter is applied inside the search, so you get a full set of results from the sources you chose rather than whatever survived a general sweep. Same question, same day:

• Talmud ticked on its own returns Kiddushin 30b — the sugya on honouring

parents — then Ketubot 103a and Yevamot 5b.

• Tanakh on its own returns Exodus chapter 20, though it still puts verse 1

above verse 12.

So if an answer is reaching past the obvious source, narrow the sources and ask again. And tell us: the questions where this happens are how we fix the ranking.

Podcasts now need a free account

Turning a synthesis into a narrated episode costs real money in text-to-speech, and until this week a signed-out visitor could do it. Asking questions and reading answers stay free and need no account. Making a podcast now needs one, and signing in is still free.

What is next

The index refreshes itself monthly from Sefaria's published export, on the 4th, picking up only what changed. Sefaria adds about nineteen English texts a month and we would rather not drift behind them again.

SeferAI is a study aid. Its syntheses are produced by a language model from the sources it retrieves, and every claim links back to the text it came from on Sefaria, so you can check it. They are not rulings, and they should be attributed to SeferAI rather than to Sefaria or to any authority quoted inside them.