How each feature works, and why it matters

Under the hood

Every source gets a health score based on freshness, retrieval hit rate and citation rate. If a source goes stale or starts causing bad answers, the system flags it and automatically reduces how much it relies on that content. Sources live inside each assistant, so what one bot knows is kept separate from another.

Under the hood

The reading-level control (Simple / Standard / Detailed) is sent to the model so replies are genuinely rewritten, short jargon-free sentences for Simple, without changing the facts, which helps SEND, ESL, low-literacy and cognitive accessibility. Text size, high contrast, a dyslexia-friendly font (readable type with extra letter, word and line spacing) and reduce-motion are applied in the widget itself and persist on-device. Read-aloud and voice input use the browser's built-in speech, so they are private and free. It all sits on top of WCAG 2.2 AA conformance, full keyboard navigation and screen-reader support across the platform and every deployed widget.

Under the hood

Confidence is the assistant's own retrieval score, shown as High, Medium or Lower with a one-line explanation, and the cited sources expand to the passages that were actually used, linking back to the original where there is a URL. The per-message Simplify calls a lightweight rewrite that keeps every fact the same but drops the reading age, complementing the global reading-level control. Together they turn a black-box chatbot into something a visitor can verify and understand.

Under the hood

A single model pass produces a summary, a sentiment read (positive, neutral, negative or frustrated), an urgency rating and a suggested next step, stored against the conversation. The inbox shows a sentiment badge and the summary on every row, and a smart-triage score blends that sentiment and urgency with the priority your team set and how recent the thread is. It's a Pro-plan feature, designed to cut the time a human spends working out which conversation to pick up next.

Under the hood

Payments run on Stripe Connect against your own connected account, so money lands with you and comxbot never touches it. A successful payment fires a payment.received event to your outbound webhooks and is recorded for your records, closing the loop between a conversation and revenue. No extra checkout page to build — the assistant offers the right product from your catalogue in context.

Under the hood

Keys are encrypted at rest and resolved in order: an assistant's own key, then the workspace key, then a comxbot-managed key. If you'd rather not open a provider account, comxbot can provide and manage a key for you on pre-paid credit, with optional automatic top-ups so the assistant never stops for lack of balance. Anthropic powers chat only, so the knowledge search uses an OpenAI or Google key.

Under the hood

Every action is typed with a schema, logged with an audit trail and optionally requires approval. The assistant proposes an action, you review it in the inbox and approve or reject it. Failed actions get retried automatically.

Under the hood

Handoff triggers include low confidence, explicit user request, failed actions, unsafe intent and repeated unanswered questions. Each handoff creates an inbox thread with priority, SLA timers and assignment.

Under the hood

The recovery ladder runs through four steps: retry retrieval with an expanded query, ask a clarifying question, offer a form or human handoff, and finally log a source gap event for your knowledge team to address.

Under the hood

Each eval suite contains test cases with expected topics, banned phrases and expected tool calls. Runs produce per-case scores and an overall pass/fail verdict. You can compare scores across versions to catch regressions.

Under the hood

Health scores are computed from four weighted factors: freshness against the SLA window, parse success rate, 7-day retrieval hit rate and 7-day citation rate. Sources scoring below 0.3 are flagged as critical and effectively excluded from retrieval.

Under the hood

Forms are schema-first. You define fields with types, labels and validation rules. The same schema renders in multiple contexts. Submissions link to conversations and can trigger webhook actions or email notifications.

Under the hood

Analytics cover conversation volume, message counts, token usage with cost estimates, handoff status breakdown, event type distribution and daily trends. Filter by period and assistant to drill into specific performance.

Under the hood

ThinkGuide uses a scaffolded prompting pipeline grounded in established pedagogical approaches: Socratic questioning to surface understanding, scaffolded instruction to build knowledge progressively, and metacognitive reflection to develop learner self-awareness. Configurable hint depth (1 to 5 levels) lets institutions control how much support the AI provides. Each step is logged with timestamps so tutors can review how the learner progressed. The system respects difficulty ceilings set per module and falls back to direct answers if the learner requests it.

Under the hood

Guardrails run as a pre-processing layer on every message. They check for assignment completion patterns, off-topic drift and content policy violations before the response reaches the learner. Blocked responses are replaced with a friendly redirect and the event is logged in the safeguarding audit trail.

Under the hood

Accessibility modes modify the system prompt and response formatting in real time. Plain English mode caps reading level at Flesch-Kincaid grade 8. Chunked mode splits responses into numbered steps with confirmation prompts. Guided mode activates the ThinkGuide pipeline. Preferences are stored per session and optionally per learner profile.

Under the hood

Setup wizard substitutes your values into the template's {{placeholders}}, creates the assistant in DRAFT status, queues starter knowledge sources for background ingestion, and writes your meeting URL, callback phone and contact email into the assistant's chat-time settings so they surface naturally in conversation. FE templates come pre-loaded with gov.uk sources (16-19 Bursary, Adult Skills Fund); SMB templates ship with UK-specific compliance language (Cadent gas, Companies House, GP NHS escalation).

Under the hood

The playground hits a separate /api/v1/playground/chat endpoint that's workspace-auth and ephemeral, no Conversation rows are saved, so QA chatter doesn't dirty your analytics. Citations are enriched with source name, mime type, similarity percentage and a 240-char preview of the retrieved chunk. Low-retrieval answers automatically log the question to your Unanswered queue.

Under the hood

The Unanswered queue surfaces frequency-ranked questions across all assistants. The 'Draft with AI' button calls a separate endpoint that retrieves relevant chunks for the question and asks the LLM to write a grounded answer with no hedging. On approve, a new KnowledgeSource of type FAQ is created and immediately re-ingested into the vector index.

Under the hood

Each knowledge chunk is embedded into a 1,536-dimensional vector with OpenAI's text-embedding-3-small model. Cosine similarity search runs through a pgvector ivfflat index in Postgres, so retrieval scales to millions of chunks per workspace without leaving your database. Health-aware re-ranking applies a multiplier based on each source's recency and citation rate before the top-k chunks are passed to the LLM.

Under the hood

OAuth flow stores per-workspace tokens. The folder picker browses Shared Drives too, with file previews, search, and a size summary so admins know what's about to be ingested before clicking. The worker walks up to 2 subfolder levels deep (capped at 50 files) and extracts text via the right parser per type: Google Docs/Sheets/Slides exported as text, PDFs via unpdf, Word via mammoth, Excel via SheetJS, PowerPoint via officeparser.

Under the hood

Connect a Twilio Account SID, Auth Token and WhatsApp number through a guided setup. Comxbot generates the webhook URL to paste back into Twilio. Inbound messages route by To-number; the assistant replies via TwiML. Staff-initiated replies from /inbox use Twilio's REST API so the response lands as a real WhatsApp message in the visitor's thread. SMS uses the same plumbing.

Under the hood

The loader script is ~3KB and lazy-loads the iframe only on first click, so visitors who never engage don't pay a perf cost. The widget renders markdown and auto-links URLs in replies, persists the conversation across page reloads via localStorage (7-day TTL), goes fullscreen on phones, animates open/close, and shows enriched citation chips with source name and chunk preview on hover.

Under the hood

Each handoff creates an InboxThread with a Handoff record carrying reason, priority, summary and suggested next step. The reply editor's Draft button calls the same playground endpoint with the visitor's last message and full history, the draft is grounded in the same retrieval the bot uses. WhatsApp/SMS replies route back through Twilio automatically; web-widget replies appear in the next page load.

Under the hood

A Vercel cron job runs once a day, builds metrics for the previous UTC day per workspace, and emails every admin via Resend. Quiet days are skipped automatically. Each email has a one-click unsubscribe link signed with HMAC-SHA256 so anyone who clicks it can opt the workspace out without logging in.

Under the hood

Each template ships with a tuned system prompt, a recommended model and temperature, suggested knowledge-source packs and an assistant kind (which controls which tools are available, so CLEARING_ADMISSIONS auto-enables the course recommender). It's UK-specific by default: £ pricing, NHS, 999 and Cadent gas escalations in trades templates, and England childcare funding bands.