The Sefaria Library Assistant is an AI-powered chatbot that helps users discover and explore Jewish texts, at any level of expertise.

Rather than providing simple summarized answers, it guides users through Sefaria's extensive library with nuanced, cited responses that reflect the depth and diversity of Jewish textual tradition — and link back to the primary sources themselves.

This blog post will explain how the Library Assistant is designed and how it was developed.

Sefaria’s Library Assistant builds on the AI assistant Claude with a sophisticated layered architecture built by Sefaria’s engineering team.

When a user’s question is submitted, a router classifier analyzes the intent and directs it to one of three specialized modes:

1. Deep Engagement – for scholarly exploration of the Library and detailed textual analysis
2. Discovery – for general learning and navigation
3. Translation – for language assistance

Each mode uses these tailored tools to answer queries:

• Semantic search finds conceptually related texts, even without matching keywords
• Keyword search locates specific terms and phrases
• Topic webs and linked texts reveal connections across the library (topic webs are thematic groupings of sources in the Library and form the underlying structure for Topics on Sefaria)

The Library Assistant only cites and links to sources in the Sefaria Library. If the Library does not support an answer, the Library Assistant is programmed to explain its limitation, not invent an answer or pull from resources outside of the Library.

As people increasingly turn to AI systems for Jewish learning, Sefaria’s Library Assistant serves as a supportive tool equipped with the built-in guardrails and grounding in primary sources that our sacred textual tradition deserves.

We’ve built the Library Assistant with multiple safeguards to maximize accuracy and reliability (though it is important to note that generative AI is not full-proof and can make mistakes).

Both human content experts and AI-powered tools monitor the chatbot's responses. If the chatbot isn't performing as expected, our team can update the guiding prompts in real time — no new code required. This means we can make immediate adjustments to improve the chatbot's behavior.

We also collect and review feedback from users through a response rating system and feedback field within the Library Assistant interface, as well as through this AI Feedback Form. This information directly informs our team on the quality of the chatbot’s responses and helps us to make improvements to the tool.

Sefaria's Library Assistant offers several features that set it apart from other tools.

1. It’s built right into the Sefaria Library.
The Library Assistant is accessible from any page of the Sefaria Library, which means you can ask questions without leaving a text you’re studying. Plus, the Library Assistant will be aware of the specific content you’re looking at and can tailor its responses to the given context.

2. The Library Assistant comes with guardrails custom built for Torah study. These are designed to verify link authenticity, keep the tool from creating independent interpretations, and prevent it from responding with a psak halakhah (ruling on Jewish law). These protections are not built into the main frontier AI models.

Additionally, the Library Assistant is programmed to reflect the diversity of perspectives of the texts in the Library, when relevant.

3. The Library Assistant draws on Sefaria's vast library of interconnected Jewish texts. This includes translations, commentaries, and cross-references, all of which help ensure its answers are grounded in verified sources from the Sefaria Library — always cited and linked.

When you use the Library Assistant, your questions are sent to our server — but they are never linked to your identity. Instead, each question is associated with a randomly generated user ID and session ID, making it impossible to connect any question to a specific person.

This means you can use the Library Assistant with confidence that your privacy is protected, while still allowing us to use aggregated data to improve the quality of the Assistant's responses over time.

>> To learn more about Sefaria’s approach to AI and the history of our tech development, visit Sefaria.org/ai.

>> Are you a software developer or simply interested in what goes on under the hood at Sefaria? Subscribe to our quarterly Developers’ Digest, where we share the latest Sefaria tech news, resources for developers, and the nuts and bolts of building Torah technology.