Some Tech Updates: The Newly Launched Gemini 3 for Hebrew OCR; Updates to ChavrutAI Chatbot; and on a General AI Browser Assistant as a Talmud Study Partner

The Newly Launched Gemini 3 for Hebrew OCR; Updates to ChavrutAI Chatbot; and on a General AI Browser Assistant as a Talmud Study Partner

Outline

The new Google Gemini 3 is really good at transcribing Hebrew manuscripts¹

ChavrutAI Chatbot: Current Status

1. What It Does

2. How It Works

3. Usage Limits

4. Interface Features

5. Query Enrichment Example

6. What’s Next

Can a General AI Browser Assistant Act as a Talmud Study Partner?

1. Full text of Comet’s response

The new Google Gemini 3 is really good at transcribing Hebrew manuscripts

Gemini 3 launched yesterday,² and it’s really good at transcribing historical manuscripts.³

EB: transcribe the hebrew handwritten text of this medieval kabbalistic manuscript

Google Gemini 3:

Here is the transcription of the handwritten Hebrew text from the image.

The script is a medieval Sephardic semi-cursive (Mashait) script, typical of Kabbalistic commentaries from the 13th–15th centuries. The text appears to be an introduction to a commentary (possibly on a work like Sefer Yetzirah or a specific mystical treatise), analyzing the author’s intent and the hidden meanings (Gematria) in the opening verses.

Hebrew Transcription:

1. דע כי כוונת המחבר היתה בזאת לקצר המקומות, יש בו שני סדרים, אחד

2. על צד החקירה והעיון ואחד על דרך הקבלה האמתית. ועוד כי במקומות רבים

3. האריך הסיפור והפרט [נהג?] בהגדת ההכרעה ואחר כך כלל אותו בסוף ספרו

4. הכתוב בקצת מקומות ובאר לך הסיפור שלו על דרך כללי כדי להקל המעיין

5. ואחר כך יבאר כל ספור וספור בפרטו בפני עצמו עד סוף הספר:

6. כוונת המחבר שכתב שני פסוקים בראש הספר והם מגדל

7. עוז וכו’ יהי לבי תמים וגו’ מנין תיבות מן שני פסוקים הם ט”ו לרמוז כי

8. הכונה הראשונה היא לקנות שערי בינה עם היראה, הפסוקים מורים

9. אל הכונה הראשונה וזהו שתחילת שמו היה שם י”ה וכבר מנו ספר יצירה כל מנין

10. [דרכי?] הזה הכל אל חיים, כמו הוא המחשבה ולקץ האמצעי וזהו כו יכונן צדיק

11. שבזה הכל תלוי והיינו סופו בתחילתו ותחילתו בסופו [?] ידע שמי י”ה, הפסוק השני

12. מתפלל לזה באמרו יהי לבי תמים בחקיך, כי סוף המעשה ועיקר הקבלה תוך

13. הלב מנוקה ועל כן אמר יהי לבי תמים באותה המדה ששמה סופו בתחלה ולשון

14. לא אבוש כענין הראשון, כי מתחלה לא יתבוששו, ובהרגל בעץ הדעת

15. הבושה והענין [באה?] וההבעה מניית תיבות אל אלו והכונה דביקה

16. אל הראייה כי דע לך כי כל הסיפור הבא למעלה מן החכמה בא לברר

Translation/Summary of Key Points:
The author explains that the work has two approaches: philosophical inquiry (”Chakira”) and true Kabbalah (”Kabbalah Amitit”).
He analyzes why the author of the book placed two specific verses at the beginning:

1. “מגדל עז...” (Migdal Oz... - Proverbs 18:10)

2. “יהי לבי תמים...” (Yehi Libi Tamim... - Psalms 119:80)

He notes that the total number of words in these two verses is 15 (ט”ו), corresponding to the Divine name Yah (י”ה) and referencing the Kabbalistic concept from Sefer Yetzirah regarding “Its end is in its beginning” (סופו בתחילתו). He connects the phrase “I will not be ashamed” (לא אבוש) to the state of Adam and Eve before the sin of the Tree of Knowledge (”לא יתבוששו”).

“””

ChavrutAI Chatbot: Current Status

The ChavrutAI chatbot provides AI-powered study assistance for Talmud texts using OpenAI’s GPT-4o model, now enhanced with a knowledge base of 671 scholarly blog posts.⁴

What It Does

When you’re studying a Talmud page, a chat button appears at the bottom of the screen. Click it to open a conversation window where you can ask questions about the text you’re reading. The AI knows which tractate, folio, and section you’re viewing and tailors responses accordingly.

The chatbot draws from two sources: the current Talmud text on your screen and a database of blog posts from ezrabrand.com covering Talmud scholarship. When you ask a question, the system searches this knowledge base for relevant content and weaves it into the response.

How It Works

Your question gets enriched with context from the current page—the tractate name, folio number, and key terms from the Hebrew and English text. This enriched query searches through 6,477 text chunks using semantic similarity, finding blog posts that discuss related topics even if they use different wording.

The AI receives your question along with the relevant Talmud section and 2-3 scholarly excerpts from the blog database. It generates responses up to 180 tokens that connect the text you’re studying with insights from the knowledge base. Citations appear as clickable links to the original blog posts.

Screenshot, from query on https://chavrusa-builder-chatbot-ezrabrand.replit.app/tractate/shabbat/31a :

EB: Explain section 2 of this Talmud page

ChavrutAI chatbot:

In Shabbat 31a, Section 2, a man persistently challenges Hillel’s demeanor by repeatedly interrupting him, even as Hillel is busy with Shabbat preparations. The man asks why the eyes of Tadmor’s residents are bleary. Demonstrating characteristic patience, Hillel responds respectfully, attributing it to the sandy environment affecting their eyes. This encounter showcases Hillel’s calm and patient nature even under pressure.

According to “Hillel Vs The Heckler A Friday Afternoon,” the narrative occurs on Erev Shabbat, adding time pressure and highlighting Hillel’s composed response amidst his preparations, further underscoring his exemplary patience.

Sources:

Hillel Vs The Heckler A Friday Afternoon

Pt1 Study Eating Divine Credit And

Tannaitic Aphorisms Avot 32 12

08:40 PM 2273 tokens

Usage Limits

You get 5 conversation turns per Talmud page and 20 turns total per day. These limits reset when you navigate to a different page or when the day changes. Conversation history persists as you move between pages, so you can continue previous discussions.

Interface Features

The chat modal sits in the bottom-right corner of the screen and can be resized by dragging from the top-left handle. Messages appear in chronological order with automatic scrolling to show new content. A counter displays remaining turns for the current session and daily usage.

When the AI is processing your query, you see a “thinking” indicator. Responses include inline citations that link to source material and a complete source list at the bottom. Quick action buttons suggest common queries like “Explain section X” or “What’s the main point here?”

Query Enrichment Example

A raw query like “Summarize section 3” becomes “Berakhot 5a tremble incite inclination Summarize section 3” by adding the tractate name, folio, and significant words from the text. This ensures the knowledge base search returns blog posts actually discussing Berakhot 5a concepts rather than generic content about summarization.

What’s Next

Potential improvements include adding hover previews for citations, displaying similarity scores to users, implementing analytics to track popular queries and identify knowledge base gaps, and experimenting with different chunk sizes and similarity thresholds for better retrieval accuracy.

Can a General AI Browser Assistant Act as a Talmud Study Partner?

Can a general AI Browser Assistant (such as Perplexity’s Comet browser AI assistant) be the better solution for fulfilling these aspirations, as opposed to a chatbot native to the website? I don’t know. I think it’s still too early to tell. The tech for both modalities is still developing at a fairly rapid pace. I plan to continue testing new tools as they progress.

Compare Joshua Waxman’s recent “Comet, Perplexed by Rabbi Eleazar” (Nov 14, 2025):⁵

Since the TechRav recently wrote in the Jewish Link about how the Atlas browser could transform Torah learning (“The Atlas Revolution: Turning Every Tab Into Torah”, Oct 30, 2025), I thought I’d see how well it understood the players in the sugya. Alas, even after force-upgrading my iMac’s operating system, I wasn’t able to run that browser because it lacked a Silicon chip. So, I instead turned to Perplexity’s Comet, which is another Agentic browser, and asked it to tell me which Sages interacted on the current page.

I was less than impressed.

See Waxman’s analysis there of that particular test of his.

Screenshot, from my actual recent interaction with Perplexity browser:

https://chavrutai.com/tractate/sotah/13a, with Perplexity browser AI assistant on the right side, after clicking on the dedicated “Summarize” button.

It’s important to note that this doesn’t solve the issue of adding the additional context needed, beyond the current page.

Full text of Comet’s response

Note: These responses were without any additional custom instructions. This is from a single convo. I copied directly from the convo using the “copy” button, and didn’t make any additional changes, other than very slight formatting changes.

“““

EB: Summarize the current webpage (=https://chavrutai.com/tractate/sotah/13a)

Perplexity Comet AI assistant: The page Sotah 13a from Talmud Bavli discusses several midrashim and teachings focusing primarily on the events surrounding the death, burial, and legacy of Jacob (Yaakov) and Joseph in the context of Jewish tradition:

• Miriam’s Prophecy: As a child, Miriam foretold the birth of Moses, who would save the Jewish people. When Moses was born, the house was filled with light, and her father acknowledged her prophecy. However, after Moses was placed in the river, her father questioned the prophecy’s fulfillment, leading to Miriam watching from afar to see what would happen to Moses.

• Jacob’s Burial: The Mishnah notes Joseph’s merit in honoring his father and arranging an extravagant burial attended by Egyptian dignitaries. The Talmud examines the shifting order of prominence between Egyptians and Joseph’s brothers at Jacob’s funeral, noting that once the Egyptians saw the Israelites’ honor, they showed more respect.

• The “Goren HaAtad” Episode: Explains that various nations, initially combative, came to Jacob’s burial but were awed by Joseph’s status, leading them to honor Jacob by hanging their crowns on his coffin—thirty-six crowns in total.

• Esau’s Objection: At the Cave of Machpelah, Esau protested Jacob’s burial, claiming the remaining grave space. Jacob’s sons rebutted, noting Esau had sold his rights; when Esau demanded proof, Naphtali ran back to Egypt for the deed. Meanwhile, Hushim the son of Dan, unaware of the details and upset at the delay, killed Esau, fulfilling Rebecca’s prophecy that she would be bereaved of both sons in one day—their burials occurred on the same day.

• Joseph’s Burial: The discussion shifts to Joseph’s greatest honor—Moses himself arranged the exodus of Joseph’s bones. The Talmud portrays Moses searching for and retrieving Joseph’s coffin from the Nile based on information from Serah bat Asher. A different view suggests Joseph was buried in a royal crypt, and Moses found his casket after speaking a declaration.

• Two Arks: While wandering in the wilderness, two arks traveled together: Joseph’s casket and the Ark of the Covenant. People wondered why a dead person would travel with the Shekhinah, to which it was explained that Joseph fulfilled the Torah’s commandments.

The page explores themes of prophecy, honor, legacy, the fulfillment of oaths, and the intertwining of Jewish historical and spiritual narratives, using both aggadic stories and halakhic insights to deepen understanding of these biblical events.​

EB: focus on the last two - “Joseph’s Burial” and “Two Arks”, what are some unusual and important hebrew terms

Perplexity Comet AI assistant: Here are some unusual and important Hebrew terms from the sections “Joseph’s Burial” and “Two Arks,” along with their explanations and significance:

• ארון (Aron)

  • Meaning: Casket or ark.

  • Usage: Refers both to Joseph’s coffin and the Ark of the Covenant—central to the “Two Arks” narrative.

  • Significance: The double meaning plays on the juxtaposition of a “dead man’s casket” and the “Ark of the Shekhinah” traveling together in the desert.

• נילוס (Nilus)

  • Usage: The Nile River.

  • Significance: Where Joseph’s casket was hidden/submerged in the Egyptian narrative. The Talmud uses the Greek/Latin loanword.

• מתכת (Matkhet)

  • Meaning: Metal.

  • Usage: Joseph’s coffin is described as being made of metal—unusual for the period and emphasizing its miraculous floating.

• סרח בת אשר (Serach bat Asher)

  • Meaning: Serach, the daughter of Asher.

  • Significance: The lone surviving witness from the generation who revealed the location of Joseph’s burial.

• קברניט (Kavranit)

  • Meaning: Royal crypt (from the Greek/Aramaic).

  • Significance: Alternate tradition where Joseph was interred among Egyptian royalty.

• שכינה (Shekhinah)

  • Meaning: Divine Presence.

  • Usage: In reference to the Ark of the Covenant, as the “Ark of the Shekhinah.”

• קיים זה כל מה שכתוב בזה

  • Meaning: “This one (Joseph) fulfilled all that is written in this (the Torah/Ark).”

  • Significance: Justifies why Joseph’s casket traveled alongside the Ark of the Covenant.

• מת (Met)

  • Meaning: The dead (referring to Joseph).

  • Usage: In the comparison between the Ark of the Covenant and Joseph’s casket.

These terms reflect both the linguistic richness and the layered symbolism of the narrative, highlighting how rabbinic tradition weaves etymology and legend to teach spiritual and ethical lessons.

“““

1. General note: this three-in-one blogpost is stylistically “rougher” than my usual ones; I didn’t fully polish it.

2. Available to test for free at https://aistudio.google.com/.

3. See the discussion a month ago by Mark Humphries, testing OCR of early modern English handwritten manuscripts in a beta version of Gemini 3, “Has Google Quietly Solved Two of AI’s Oldest Problems?” (Oct 17, 2025).

Compare this previous AI OCR test of mine on this same medieval kabblistic manuscript (Yeriah Gedola), two years ago, see there for my own transcription of the initial lines: “General Update, and Discussion of Widely Available Cutting-Edge Software Tools that Read Text in Images (=OCR), including Hebrew” (Oct 24, 2023).

I’ve been consistently testing since then, and Gemini 3 is by far the best of any that I’ve ever tested; it’s a huge jump.

And compare also my related “OCR Odyssey: Taming Text Recognition for Hebrew Classics” (Nov 05, 2023)

4. The app: https://chavrusa-builder-chatbot-ezrabrand.replit.app/.

It is a fork of an earlier iteration of my main ChavrutAI website.

Registration is required to use the chatbot.

See my previous discussion of the chatbot here: “Developing the ChavrutAI Chatbot, an AI Talmud Study Partner” (Aug 14, 2025).

Note that the enriched chat is very much a work-in-progress. On a technical level, it uses Retrieval-augmented generation - RAG.

Technical Implementation:

The system uses text-embedding-3-small to convert text into 1,536-dimensional vectors stored in a PostgreSQL database with the pgvector extension. Semantic search runs in under 50 milliseconds using cosine similarity. An in-memory cache stores query embeddings for one hour, reducing API costs by approximately 70%.

Blog posts were processed from PDF files, chunked into 500-token segments with 50-token overlap to preserve context, and embedded. The database contains 671 posts totaling roughly 3 million tokens of content.

5. Note: I’ve been testing Perplexity Comet for this purpose for over a month, before either of these articles came out.