Ask. Read. Follow up.
A React interface and .NET backend connect the conversation to a source reader with passage, edition, and attribution.
Ask
Answers you can trace.
Jewish questions, grounded in Torah, Mishnah, Talmud, and other primary texts.
Visit askarabbi.aiBuilt end to end by Amitai · External AI models
Follow the answer.
Recorded example · Shortened excerpt, not a live chat
The system around the models
Grounding is a product decision.
The model is one part of the application. I built the retrieval, evidence boundaries, validation, and reading experience around it.
Bound the evidence.
Corpus retrieval and manifest checks prepare a source packet. Structured answers refer to its evidence identifiers.
Make it a product.
Authentication, conversation storage, and container deployment support the end-to-end application.
Production architecture: React · .NET · Azure OpenAI retrieval · Cosmos DB · Azure Container Apps. The AI models are external.
Process walkthrough Step 1 of 6
Start with the question.
The application accepts the question and conversation context, then prepares a grounded-answer request.
Search the approved corpus.
Retrieve relevant passages from the approved text collection. Corpus and manifest checks constrain which sources can enter the answer.
Give the model bounded evidence.
Build a limited evidence packet with stable source identifiers. The model works from that packet instead of inventing citation metadata.
Draft against the sources.
An external AI model proposes a structured answer, linking its claims and quotations to the supplied evidence.
Checking the answer.
Exact quotation checks and a separate claim-support review before an answer is shown.
Keep the evidence within reach.
Show the checked answer with citations and a source reader. Readers can inspect the passage, edition, and attribution themselves.
What if a check fails?
The system permits one repair using the same evidence packet, then checks again. If the evidence is insufficient or validation still fails, it returns a failure state instead of presenting that draft as a grounded answer. These checks reduce risk; they do not guarantee a correct interpretation.
What I built.
The application, retrieval pipeline, citation checks, source reader, and infrastructure around the models.