Knowledge > Products > Chatbot
Chatbot Overview
What It Is
The ChurchWiseAI Chatbot is an AI-powered ministry tool for churches -- not a generic question-answering bot. It is deployed as a care agent that represents the church to every person who chats with it, building relationships on behalf of the pastoral team, connecting visitors to ministries, capturing prayer requests, scheduling pastoral appointments, and surfacing church-specific content through a theologically-grounded lens.
A church subscribes to a chatbot plan, and the chatbot immediately knows the church's name, denomination, staff, hours, ministries, events, theological tradition, and doctrinal positions. It speaks from within that tradition, not about it.
Pastoral Model (Not ChatGPT Model)
The chatbot follows the HEAR protocol, not a search-engine pattern:
- Hear -- Read between the lines. "What time are services?" really means "I'm nervous about visiting." "Are you affirming?" really means "Will I be accepted here?"
- Empathize -- Acknowledge the feeling behind the question before answering the fact.
- Connect -- Link their need to something specific at the church: a group, a ministry, a staff member, an event.
- Invite -- Proactively offer a next step: "Would you like me to have someone from our team reach out?"
This is fundamentally different from a ChatGPT-style chatbot. The system prompt enforces strict length limits (max 2 short paragraphs), one call-to-action per response, one follow-up question per response, and empathetic hearing before factual answers. It uses progressive disclosure -- earning trust before asking for contact information, never cold-asking.
ONE API Endpoint
All chatbot traffic flows through a single endpoint:
POST /api/chatbot/stream
This endpoint serves all properties (churchwiseai.com, pewsearch.com Pro Websites, embed widgets) and all chatbot tiers. It routes internally to three chatbot types based on plan and source:
| Chatbot Type | Source | Plan | Tools | Prompt Depth |
|---|---|---|---|---|
| Pro Website | PewSearch Pro Website bundle | pro_website | 1 (prayer request) | Church-info focused, upsell to full |
| Full | ChurchWiseAI subscription | starter / pro / suite | 12-39 (by tier) | Full HEAR protocol, agentic tools |
Three UX Surfaces
The chatbot is accessible through three interfaces, all hitting the same API:
| Surface | URL Pattern | Use Case |
|---|---|---|
| Care Hub | /care/[slug] | Full care experience with persona selection (grief, welcome, youth, etc.) |
| Hosted Chat Page | /chat/[slug] | Standalone chatbot page, embeddable link |
| Embed Widget | JavaScript snippet on any domain | Floating chat widget on church websites |
Four Agents
The chatbot uses a four-agent architecture: Coordinator, Care, Stewardship, and Discipleship.
| Agent | Role | Tool Count |
|---|---|---|
| Coordinator | Front door for all chats. Handles general inquiries, service times, directions, visitor intake, events, children's ministry, volunteer sign-up, staff messaging. Routes to specialized agents when needed. | 14 tools |
| Care | Pastoral care specialist. Prayer requests, grief support, crisis intervention, pastoral visits, callbacks, counseling scheduling, safety flagging, local resources. | 10 tools |
| Stewardship | Financial ministry. Giving links, giving history, benevolence requests. Handles generosity naturally -- never cold, never guilt-inducing. | 3 tools |
| Discipleship | Spiritual growth. Bible verse lookup, small group finder, sermon archives, devotionals, theological deep dives, lesson plans, illustration search, worship playlists. | 12 tools |
Why Four Agents in Chat, Two in Voice
The voice agent consolidated to two agents (Coordinator + Care) because phone calls are linear and every handoff creates an awkward silence gap. Chat has no such constraint -- text-based conversations tolerate agent specialization without UX friction. The chatbot routes tool calls to the appropriate agent via the TOOL_AGENT_MAP without visible handoffs.
39 Tools Across 3 Tiers
Tools are gated by subscription tier. Higher tiers include all tools from lower tiers:
| Tier | Tools Available | Monthly | Annual |
|---|---|---|---|
| Starter | 12 tools (visitor care, prayer, callbacks, directions, sermon info, announcements, Bible lookup, connection card, giving link, event registration, safety flag) | $14.95 | $149.50/yr |
| Pro | 35 tools (adds small groups, volunteering, pastoral visits, care needs, grief resources, follow-up, counseling, devotionals, facility booking, sermon archive, worship playlist, kids info, benevolence, child check-in, subscriptions, staff messaging, local resources, illustrations, theological deep dive, lesson plans, devotional generation, conversation summary) | $34.95 | $349.50/yr |
| Suite | 39 tools (adds giving history, engagement monitoring, weekly reports, draft follow-up messages) | $59.95 | $599.50/yr |
All chat plans include a 14-day free trial and support both monthly and annual billing (annual saves ~17%).
LLM Strategy
| Role | Model | When Used |
|---|---|---|
| Primary | Claude Haiku 4.5 | Default for all conversations |
| Escalation | Claude Sonnet 4.6 | Complex/sensitive conversations (auto-detected) |
| Fallback | OpenAI gpt-4o-mini | Automatic failover if Anthropic returns 5xx or times out |
Prompt caching is enabled on system prompts (cache_control: ephemeral) for ~60% input cost reduction on Anthropic.
RAG System
The chatbot draws from three knowledge sources, retrieved in parallel:
| Source | Table | Records | Purpose |
|---|---|---|---|
| Theological RAG | unified_rag_content | 327K+ | Sermon illustrations, devotional content, theological resources -- filtered by the church's theological lens |
| Church Knowledge Base | church_knowledge_base | Per-church | Church-specific FAQs, uploaded documents, custom content (managed via admin dashboard) |
| Product Knowledge | product_knowledge | Shared | PewSearch/ChurchWiseAI feature info, billing, troubleshooting -- injected into all chatbot instances |
Embeddings are generated via OpenAI text-embedding-3-small (1536 dimensions). The RAG query returns the top 8 theological matches and all relevant church KB matches, formatted as context blocks in the system prompt.
Theological Lens System
Every chatbot response is grounded in the church's theological tradition. The system supports 17 traditions plus a universal Christocentric default.
Resolution priority:
- Client-side override (demo mode only, validated against known lens IDs 1-17)
- Church-specific setting in
church_theological_lensestable - Auto-detection from denomination via
DENOMINATION_TO_LENSmapping - Default: Christocentric (lens ID 10)
The lens drives three layers of theological grounding:
- Doctrinal rules from
theological_contradictions-- must-include/must-exclude terms per doctrine category (e.g., Baptist lens enforces credobaptism, excludes infant baptism) - Lens vocabulary from
lens_knowledge-- preferred and avoided terminology for natural-sounding responses - Church-specific overrides from
organization_settings.doctrinal_overrides-- custom practices that override standard positions (e.g., a Reformed church that also does dedications)
Moderation System
The chatbot enforces a four-level escalation ladder:
| Level | Trigger | Response |
|---|---|---|
| Level 1 | Inappropriate/creepy comments | Firm gracious boundary, one redirect |
| Level 2 | Profanity, insults, verbal abuse | Acknowledge without absorbing, boundary after 1 redirect, disengage after 2 |
| Level 3 | Threats of violence | Immediate safety flag (critical), crisis resources, stop engaging |
| Level 4 | Self-harm / suicidal ideation | Mandatory 988 + 741741 + 911 resources, urgent safety flag, pastoral connection |
Progressive restrictions enforce consequences:
- 2 violations: 5-minute cooldown
- 4 violations: 24-hour temporary block
- Severe/repeated: permanent block
Crisis safety net: A regex-based pattern matcher runs as a non-negotiable post-processing step. If the user message matches self-harm/crisis patterns and the LLM response is missing any of the three mandatory resources (988, 741741, 911), they are auto-appended. This catches LLM omissions regardless of model or provider.
Pricing
| Plan | Monthly | Annual | Tools | Trial |
|---|---|---|---|---|
| Chat Starter | $14.95 | $149.50/yr | 12 | 14-day free |
| Chat Pro | $34.95 | $349.50/yr | 35 | 14-day free |
| Chat Suite | $59.95 | $599.50/yr | 39 | 14-day free |
Bundle plans (Voice + Chat) are also available -- see PRICING.md.
Current State
- Fully deployed on Vercel via
churchwiseai-web - 3 demo churches configured (Grace Community, St. Joseph Catholic, The Bridge Community)
- 3 Stripe subscriptions — Melvindale CoG is the designated internal test account (100% coupon
cwa-validation-100pct,sub_1TIUbGFaoK5IPzNowJrWhgDM, all invoices net $0); real paying customers: Zewdei/Medhanialem (starter, 100% coupon through June 2026), Hope Community (starter, cancelling) - LLM provider layer operational (Anthropic primary, OpenAI fallback)
- Prompt caching active
- Moderation system enforced
- FAQ short-circuit operational (zero LLM cost for exact matches)
- Usage tracking and per-tier conversation limits enforced
See Also
- Architecture -- full request pipeline, two-call LLM pattern, error handling
- Tools -- complete tool catalog with schemas and implementations
- Agents -- agent specialization, persona templates, domain rulesets
- UX Model -- Care Hub, hosted chat, embed widget
- Tier Restrictions -- tool gating, usage limits, plan normalization
- Moderation -- violation types, escalation, restrictions, crisis safety net
- Widget Config -- embed snippet, origin validation, branding