Chatbot-as-a-Service¶
Service ownership
Owner: application-services (apps-pm@clouddigit.ai) — Status: GA — Last audited: 2026-05-11
Multi-channel chatbots — web widget, WhatsApp Business, Facebook Messenger, IVR — backed by Cloud Digit-hosted LLMs.
What it is¶
A turnkey chatbot platform that pulls together:
- A flow builder for conversation design
- LLM integration via LLM-as-a-Service (sovereign-resident inference)
- A knowledge-base ingestor (PDF, docx, web)
- A vector store (via Vector Database)
- Channel adapters (web, WhatsApp, Messenger, Telegram, voice)
- Human-handoff to your existing CRM / helpdesk
- Analytics dashboard
Why this exists separately from the inference layer¶
Building a production chatbot from raw inference is real work — RAG pipeline, intent classifier, fallback routing, escalation gating, audit trail, channel adapters, NLP for Bengali and Banglish, abuse filters. Chatbot-as-a-Service bundles all of it.
Use cases¶
- Banking customer support (balance enquiry, branch locator, FAQ)
- Government citizen services (forms, statuses, eligibility checks)
- E-commerce (order status, returns, recommendations)
- Healthcare (appointment booking, FAQ, triage front-line)
Languages¶
- Bengali — first-class (intent recognition, response generation)
- English — first-class
- Banglish (Bengali in Latin script) — handled
- Other languages on request
Compliance¶
- All conversations stay on Cloud Digit infrastructure
- Per-conversation retention configurable; default 90 days
- Optional zero-retention mode (no message ever stored beyond the session)
Pricing¶
Per-conversation-month + per-channel-month + the underlying LLMaaS token usage. See Pricing.
Related¶
- LLM-as-a-Service
- Vector Database
- Agentic AI-as-a-Service — for tool-calling agents (the next step beyond chat)
Operate this service¶
Hosted conversational AI — LLM-backed, with retrieval, tools, and BD-localized integrations.
Use cases¶
- Customer support (banking, telco, e-commerce)
- Internal knowledge bot (HR, IT support)
- Sales assistant (lead qualification)
- Bangla-language native chatbots
IAM¶
| Role | Can do |
|---|---|
chatbot.viewer | View bots, conversation logs |
chatbot.editor | Edit bot instructions, knowledge base |
chatbot.deployer | Deploy bots to production channels |
chatbot.admin | Manage bots, integrations, model selection |
Architecture¶
User → Channel adapter → Chatbot engine → LLM + Retrieval + Tools ↓ [Vector DB / KB] [External APIs]
Channels supported¶
- WhatsApp Business API
- Facebook Messenger
- Web widget
- SMS (via local operators)
- bKash chat
- Custom (REST API)
Knowledge base¶
Document ingestion: - PDF, DOCX, HTML, Markdown - Auto-chunking and embedding - Stored in Vector DB - Re-index nightly
Compliance¶
For regulated industries (banking, healthcare): - All conversations logged - PII detection + redaction - Audit-trail for AI decisions - Human escalation path
Related¶
Metrics¶
| Metric | Healthy | Alert |
|---|---|---|
chatbot.conversations_per_hour | varies | |
chatbot.first_response_time_ms p95 | < 2000 | > 5000 |
chatbot.completion_rate | > 80% | < 60% (users abandoning) |
chatbot.handoff_rate | < 20% | > 40% (bot weak) |
chatbot.satisfaction | > ⅘ | < 3.5/5 |
chatbot.errors_per_hour | low | spikes |
Tuning the bot¶
Weekly review: - Top 10 conversation flows (where do users go?) - Top 10 handoff reasons (where does bot fail?) - Top 10 dissatisfaction signals
Refine: - Improve knowledge base where retrieval misses - Add explicit tool calls for actions - Update system prompt for tone / boundaries
A/B testing¶
Multiple bot versions live; route by user cohort:
bash cd chatbot deploy --bot acme-support \ --traffic 'v1.4=80,v1.5=20'
Metrics split by version; promote winner.
Tool calls¶
Chatbot can invoke tools: - Look up customer order status - Check account balance - Schedule callback - Create support ticket
Each tool is a function definition + endpoint; chatbot LLM decides when to call.
Multi-language¶
Bangla + English on the same bot: - Auto-detect user language - Respond in same language - KB in both languages (or one with translation layer)
Conversation logging¶
Per channel + privacy policy: - Logged: all turns, metadata (channel, timestamp, etc.) - Redacted: PII (auto-detected; can configure aggressiveness) - Retention: 90 days default; longer for compliance
Related¶
Bot hallucinating¶
Bot invents facts not in knowledge base: - Strengthen system prompt: "answer only from provided context, otherwise escalate" - Add explicit grounding via retrieval - Lower temperature - Improve KB coverage
No bot is hallucination-free; mitigate.
Bot doesn't understand Bangla¶
Possible causes: - Wrong model selected (use bangla-llm-7b or a multilingual one) - KB only in English; user query in Bangla finds nothing - Mixed script (Bangla in Latin chars) confuses retrieval
Handoff doesn't reach human¶
User asks for human; transfer fails: - Human team unavailable (off-hours) - Channel doesn't support handoff (WhatsApp policy) - Bug in escalation workflow
For 24/7 needs: hire shifts or accept after-hours queue.
Tool call failed¶
Bot tried to invoke a tool; error: - Backend API down - Authentication issue - Schema mismatch (bot sent wrong format)
Bot should gracefully degrade — say "I can't do that right now, let me get a human." Don't expose backend errors.
Conversation history confused¶
Bot loses context mid-conversation: - Token limit reached for the model - Conversation reset by channel timeout - Bug in context management
Most LLMs handle 4k-128k tokens; long conversations need summarization.
Privacy / PII leakage¶
Bot replied with PII it shouldn't have: - KB ingestion didn't redact PII (re-ingest with stricter redaction) - Bot recovered PII from history of another user (cross-user leakage; bug) - User intentionally probed; bot disclosed
Audit; investigate; tighten.
Bot deploy broke production¶
A new bot version causes regressions: - Roll back: cd chatbot rollback --bot acme-support - Investigate in sandbox - Re-deploy with fix + A/B test before full
Cost spike¶
LLM token costs ramping: - Long context windows (more history per turn) - Bot generating excessively long responses - Many low-value conversations (broaden funneling earlier)
Cap with hard quotas; tune for efficiency.