Torah Study AI

in progress

AI-powered Torah study companion. RAG pipeline on the entire Sefaria library (3.5M texts) with verifiable sources and semantic search.

PythonFastAPILangChainWeaviateReactClaude APICohereGemini Embeddings

Your AI study partner for Torah. Real sources. Every answer.

MVP Features

#	Feature	Description
1	Chat with sources	Ask a question, get an answer with clickable Sefaria links
2	Conversation history	Sidebar with past chats, resume any conversation
3	Multi-language	French interface, Hebrew quotes, answer in user's language
7	Parashah of the week	Auto-suggest weekly Torah portion with key insights

Architecture

The system has 4 layers:

React frontend - Chat interface with shadcn/ui components and streaming responses
FastAPI backend - REST API with JWT auth (MFA email), SQLite database, SSE streaming
LangChain RAG engine - Retrieval (Weaviate) then Rerank (Cohere) then Generation (Claude)
Weaviate vector DB - 3.5M Sefaria texts indexed with hybrid search (vector + BM25)

4 Blocks

Block	What it builds	Status
1. Chat with Claude	Working chatbot with React frontend and prompt engineering	To do
2. Make it a real app	FastAPI backend, auth, persistence, Docker, deploy on Elestio	To do
3. Add Sefaria knowledge	RAG pipeline on 3.5M texts with Gemini embeddings and hybrid search	To do
4. Make it reliable	LangFuse observability, Ragas evaluation	To do

Key Constraint

Never invent a Torah source. If the RAG doesn't find a relevant text, the system says "I didn't find a source" instead of making one up. On sacred texts, hallucination is not just an error - it's an offense.

The Problem

Millions of Jews want to study sacred texts but can't. Not because they don't want to. Because:

No time. Parents, jobs, 10 minutes a day max.
No level. No Jewish school, no yeshiva. Ancient Hebrew is a wall.
No map. Torah? Talmud? Halakha? It's an ocean with no guide.
No study partner. Jewish study is meant to be done in pairs (chavruta). Alone, it's discouraging.

The tools that exist (Sefaria, Chabad.org) are libraries. It's like giving an encyclopedia to someone who doesn't know how to read.

Who is this for

This project is NOT for yeshiva students or rabbis. It's for:

Persona	Who they are	What they need
Jonathan	29, no Jewish school, lives in Israel, average Hebrew	A patient guide that explains simply in French
Sarah	32, moved to Israel 2 years ago, observant	French texts with Hebrew side by side
Michael	45, goes to synagogue on Shabbat	Understand what he reads at the synagogue

Target market: 3-5 million people. No AI tool targets them today.

What exists today

Tool	What it does	Why it's not enough
Sefaria	3.5M texts, API, open source	Static library. No guide. Keyword search only.
OHRBIT	AI Torah tool (Feb 2026)	Validates demand. Limited details.
Chabad.org	Biggest Jewish website	Massive content. Zero AI.
JewPT, Rebbe.io	AI attempts	Basic wrappers. No real RAG. No traction.

The gap

Nobody has built a real RAG on the Sefaria dataset. The opportunity:

MVP Features

#	Feature	Description
1	Chat with sources	Ask a question, get an answer with clickable Sefaria links
2	Conversation history	Sidebar with past chats, resume any conversation
3	Multi-language	Ask in French, get answer in French with Hebrew quotes
7	Parashah of the week	Auto-suggest weekly Torah portion with key insights

Our strategy

We build in 4 blocks. Each block adds a layer of intelligence:

Risks

Risk	How we handle it
Hallucination = religious offense	Strict RAG, verifiable sources, "consult your Rabbi" disclaimer
"AI Rabbi" positioning	We're a chavruta (study partner), never a religious authority
Hard to monetize	Freemium + Jewish foundation philanthropy

Skills you learn

System Architecture

Stack overview

Component	Choice	Cost
Frontend	React + Vite + shadcn/ui	$0
State management	useState	$0
Backend	FastAPI (Python)	$0
Database	SQLite	$0
Auth	JWT + MFA (email code)	$0
RAG framework	LangChain	$0
Embeddings	Gemini Embedding 001	~$2.50 one-shot
Vector DB	Weaviate (Elestio)	~$6/month
LLM	Claude Sonnet (API)	~$5-15/month
Reranking	Cohere Rerank	$0 (free tier)
Observability	LangFuse (Elestio)	$0
Parashah data	Sefaria API `/calendars`	$0
Hosting	Elestio	~$10/month
Total POC		~$20-30/month

Embeddings

Sefaria tested 18 models on their own texts and published the results:

Model	Recall@1	Cost for all Sefaria	Type
Gemini Embedding 001	93.9%	~$2.50	API
OpenAI text-embedding-3-large	69.9%	~$55	API
OpenAI text-embedding-3-small	~65%	~$8	API
DictaBERT (Hebrew specialized)	1.7%	$0	Local

Our choice: Gemini Embedding 001. 3x cheaper AND 40% more accurate than OpenAI on Rabbinic texts.

Database and Auth

Database: SQLite. A single file on disk. No server to manage. Good enough for a POC. We move to PostgreSQL when we need multiple concurrent users.

Auth: JWT + MFA with email code.

User registers with email
Receives a verification code by email (MFA)
Gets a JWT token after verification
Token sent with every API request
Stateless auth, no session storage needed

RAG Pipeline

Parashah of the week

Uses Sefaria's /calendars API endpoint which returns the current week's Torah portion. We display it as a suggested study topic when the user opens the app.

What Sefaria gives us (HuggingFace)

Resource	What it is	Impact
`hebrew_library`	3.55M text segments, already chunked	Skip manual chunking
`english_library`	886K English segments	Ready to embed directly
`links`	3.74M cross-references	Knowledge graph for V2
NER models	Detect citations in text	Auto-extract Sefaria refs
Embedding Leaderboard	18 models benchmarked	Proved Gemini is best

Block 1 - Chat with Claude

Build a working Torah chatbot that talks to Claude directly.

Step 1: First LLM call `[TODO]`

Python script that sends a Torah question to Claude and prints the answer. Learn the Anthropic SDK, system prompts, and streaming.

Step 2: React frontend `[TODO]`

Chat interface with React + Vite + shadcn/ui. Streaming display (word by word). useState for conversation state.

Step 3: Prompt engineering `[TODO]`

System prompt that turns Claude into a chavruta (study partner). Guardrails: never invent sources, never give halakhic rulings, always cite references. Multi-language: answer in the user's language.

Block 2 - Make it a real app

Add a backend, auth, persistence, and ship it live.

Step 4: FastAPI backend `[TODO]`

REST API between frontend and Claude. POST /chat endpoint with SSE streaming.

Step 5: Auth + Save conversations `[TODO]`

JWT authentication with MFA (email verification code)
SQLite database with users, sessions, and messages tables
Conversations tied to authenticated user

Step 6: History in frontend `[TODO]`

Sidebar with past conversations. Click to resume. "New conversation" button. Protected by auth (login required).

Step 7: Docker `[TODO]`

Package everything in containers. One docker compose up to run the full app.

Step 8: Deploy to Elestio `[TODO]`

Deploy Docker stack on Elestio. HTTPS, custom domain.

Block 3 - Add Sefaria knowledge

Build the RAG pipeline. Answers based on real texts.

Step 9: Load Sefaria datasets `[TODO]`

Load Sefaria/hebrew_library and Sefaria/english_library from HuggingFace. Already pre-chunked with metadata. Explore the data.

Step 10: NER enrichment `[TODO]`

Run Sefaria's NER models (en_torah_ner, he_ref_ner) on the chunks to extract citations and entities as structured metadata.

Step 11: Embeddings + Weaviate `[TODO]`

Embed all chunks with Gemini Embedding 001 (~$2.50 total). Index in Weaviate with hybrid search (vector + BM25).

Step 12: RAG pipeline with LangChain `[TODO]`

Connect Weaviate search + Cohere Rerank + Claude generation in a working pipeline.

Step 13: Advanced search + Parashah `[TODO]`

Hybrid search, metadata filtering ("search only in Talmud"). Parashah of the week feature using Sefaria API /calendars.

Block 4 - Make it reliable

Track quality and ship the final product.

Step 14: LangFuse observability `[TODO]`

Trace every request: search results, reranking, generation. See exactly why an answer is good or bad.

Step 15: Ragas evaluation `[TODO]`

Measure RAG quality on 50 test questions. Metrics: faithfulness, relevancy, context precision. Use Rabbinic-Hebrew-English-Pairs dataset for evaluation.

Step 16: Integration + deploy `[TODO]`

Connect RAG to the chatbot. Deploy the full stack on Elestio. Every answer now has verifiable Sefaria sources with clickable links.

MVP Features

Architecture

4 Blocks

Key Constraint

The Problem

Who is this for

What exists today

The gap

MVP Features

Our strategy

Risks

Skills you learn

System Architecture

Stack overview

Embeddings

Database and Auth

RAG Pipeline

Parashah of the week

What Sefaria gives us (HuggingFace)

Block 1 - Chat with Claude

Step 1: First LLM call [TODO]

Step 2: React frontend [TODO]

Step 3: Prompt engineering [TODO]

Block 2 - Make it a real app

Step 4: FastAPI backend [TODO]

Step 5: Auth + Save conversations [TODO]

Step 6: History in frontend [TODO]

Step 7: Docker [TODO]

Step 8: Deploy to Elestio [TODO]

Block 3 - Add Sefaria knowledge

Step 9: Load Sefaria datasets [TODO]

Step 10: NER enrichment [TODO]

Step 11: Embeddings + Weaviate [TODO]

Step 12: RAG pipeline with LangChain [TODO]

Step 13: Advanced search + Parashah [TODO]

Block 4 - Make it reliable

Step 14: LangFuse observability [TODO]

Step 15: Ragas evaluation [TODO]

Step 16: Integration + deploy [TODO]

Timeline

Block 1 - Chat with Claude (1 week)

Block 2 - Make it a real app (2 weeks)

Block 3 - Add Sefaria knowledge (3 weeks)

Block 4 - Make it reliable (2 weeks)

V2 - After POC

Step 1: First LLM call `[TODO]`

Step 2: React frontend `[TODO]`

Step 3: Prompt engineering `[TODO]`

Step 4: FastAPI backend `[TODO]`

Step 5: Auth + Save conversations `[TODO]`

Step 6: History in frontend `[TODO]`

Step 7: Docker `[TODO]`

Step 8: Deploy to Elestio `[TODO]`

Step 9: Load Sefaria datasets `[TODO]`

Step 10: NER enrichment `[TODO]`

Step 11: Embeddings + Weaviate `[TODO]`

Step 12: RAG pipeline with LangChain `[TODO]`

Step 13: Advanced search + Parashah `[TODO]`

Step 14: LangFuse observability `[TODO]`

Step 15: Ragas evaluation `[TODO]`

Step 16: Integration + deploy `[TODO]`