Sermon Prep RAG Assistant
A retrieval assistant grounded only on your church's own doctrine and trusted commentary. Cite-or-refuse.
A retrieval assistant grounded only on your church's own doctrine and trusted commentary. Cite-or-refuse — the antidote to invented citations.
| Category | Formation |
| Church-health domain | Send Out the Equipped |
| Data-privacy tier | Low (content is doctrinal) — but commentary licensing matters |
| Mastery-ladder target | Level 5 · Systems |
| Build stack | NotebookLM or Claude Projects (no-code start) → custom retrieval app behind Workspace SSO; query logging |
The problem
Off-the-shelf chatbots fabricate sources. The most rigorous evidence comes from law: a public tracker has logged hundreds of court decisions involving AI-hallucinated citations — roughly 90% of them in 2025 — and courts have repeatedly sanctioned attorneys (a $59,500 award in Illinois; $3,000 each against two attorneys in Colorado). Critically, the hallucinations include real quotes from real cases that fail to support or directly contradict the claim. That is precisely the danger in sermon prep: a fabricated commentary quote attributed to a real theologian, preached as truth.
What good looks like
A retrieval assistant grounded only on the church's vetted corpus — doctrinal statements, the sermon archive, and approved commentaries — with a system prompt that says "answer only with citations from indexed sources; refuse if the answer isn't in the corpus." The preaching team queries it behind Workspace SSO; every query is logged. It turns an open-ended hallucination risk into a closed, verifiable one. This is the workshop's "cite-or-refuse" discipline built into a tool.
Market scan
| Tool | Fit | Pricing (verify) |
|---|---|---|
| Logos Bible Software (+ AI) | Smart Search footnotes results to your library; labels AI content | ~$9.99–14.99/mo |
| Pulpit AI (Subsplash) | Sermon-to-content repurposing — not a grounded research tool | $39/$59/$129/mo |
| Sermonary | Sermon builder w/ integrated commentaries | ~$29/mo |
| NotebookLM (Google) | Source-grounded RAG, inline citations to your docs | Free tier |
| Claude Projects / custom GPTs | Upload corpus + instruct cite-or-refuse | Subscription |
The split: ungrounded general chatbots invent citations; grounded tools (Logos Smart Search, NotebookLM, a custom RAG) cite real, retrieved sources — but none eliminate verification. Independent testing reports NotebookLM around 13% response-level hallucination versus ~40% for ungrounded LLMs: much better, not zero.
The gap: this is a legitimate engineering pattern. Commercial tools index their libraries; your differentiator is grounding on the church's own approved corpus plus a hard refusal when an answer isn't in it.
Data privacy & security
- Low data-sensitivity — a doctrinal/sermon corpus is far less sensitive than minutes or giving.
- The real concern is IP/licensing: many commentaries are copyrighted, so only index materials you own or are licensed to use.
- Restrict the chat UI to the preaching team behind Google Workspace SSO; least privilege; log every query (good for audit and spotting misuse).
- Bake in human verification. Given that even warned professionals keep citing fake material, every quoted citation must be clicked through and confirmed before it reaches the pulpit.
How to build it
- Start no-code: load the corpus into NotebookLM or a Claude Project; set the instruction to "answer only with citations from indexed sources; refuse if not in corpus."
- Pilot with the preaching team.
- If you outgrow it, graduate to a custom retrieval app (vector index + chat UI behind Workspace SSO) with the same cite-or-refuse prompt and query logging.
- Make the human verification step a standing rule, not an afterthought.
Rollout plan
- Q1: Assemble and clean the corpus (doctrine, sermon archive, licensed commentaries). Pilot in NotebookLM.
- Q2: Formalize the cite-or-refuse prompt and the verification rule with the preaching team.
- Q3: If usage justifies it, build the SSO-gated custom app with logging.
Effort & cost
- No-code pilot: hours; near-zero marginal cost (NotebookLM/Claude tiers).
- Custom RAG app: a developer engagement (weeks) plus modest hosting and per-query model costs.
Sources
- AI hallucinated-citations tracker — https://www.damiencharlotin.com/hallucinations/
- Why lawyers keep citing fake cases — https://www.scientificamerican.com/article/why-lawyers-keep-citing-fake-cases-invented-by-ai/
- How Logos uses AI — https://support.logos.com/hc/en-us/articles/35181728416397-How-Logos-uses-AI
- NotebookLM (source-grounded RAG) — https://www.emergentmind.com/topics/notebooklm
- Pulpit AI — https://www.subsplash.com/product/pulpit-ai
Honesty flag: legal hallucination counts grow weekly ("hundreds and rising"); the NotebookLM ~13% vs ~40% figure is from one testing context — directionally supported, not a universal benchmark. Logos/Pulpit pricing from secondary sources; verify.