Every feature, from the dashboards it lives in.
What the assistant does for customers, what your dashboard gives the people who run the store, and what our own team can and cannot see.
For customers
Grounded policy answers
Returns, shipping, warranty, cancellation — answered from the documents you published, with citations. Expired policies stop being retrievable on their end date.
Product discovery
Natural-language search over the published catalogue with colour, size, category and budget understood. Cards, not paragraphs.
Live price and stock
Read at answer time in the customer's market and currency. If the live source is down the assistant says so rather than quoting an indexed number.
Order and return help
Signed-in customers can ask about their own order, shipment and whether a line is returnable — with the policy that decides it cited.
Cart with confirmation
“Add it to my cart” shows a preview and a single-use confirmation. The confirmation carries the arguments, so what was approved is what happens.
Hand-off to a person
A redacted summary and the evidence handles go to your support queue, so the person does not start from zero.
Clarifies instead of guessing
“Add it” with no item named gets a question, not a guess. Ambiguity is a first-class outcome.
Locale-aware
Sessions carry a locale and currency; prices, dates and units are formatted for the customer's market.
In your dashboard
Sign in at chatbot.trooply.ai/dashboard with your email and password.
| Page | What you see | What you can do |
|---|---|---|
| Overview | Whether your assistant is answering customers, conversations and how they were handled, the most shown products and most cited policies, questions it couldn't answer, and what needs attention | — |
| Setup | Six steps from a new account to an assistant answering on your website | Connect a Shopify store, accept how products are read, start the first sync, add a policy |
| Conversations | Every question customers asked and how the assistant answered, including the chats that still need a person | Mark a conversation as handled, or reopen it |
| Catalogue | Your store connection, sync history, product fields and the items that couldn't be read | Sync now; accept product fields or put new ones in use |
| Policies | Every policy with its status, full text and versions, and the questions your policies answered differently | Add a policy, edit its text as a new draft, publish, retire |
| Assistant | The code for your website, a private preview, and what the assistant may do | Turn a capability on or off |
| Plan & usage | This month's conversations against your plan, daily and monthly use, seats, stores and invoices | — |
| Team | Who can sign in, at which access level, and when they last did | Add a teammate, change their access, reset a password, suspend or restore |
| Store settings | Store name, time zone, language, currency and website address | Edit the details; change the website address |
| Advanced | Activity, sign-ins, privacy, quality checks, capabilities and roles, and product field versions, for technical staff | Sign someone out |
Access comes in four levels — Owner, Manager, Support and Viewer — and a page appears only for the people whose access includes it. A change asks for a note for the activity log, and a change that takes something away or changes who can sign in asks for your password again. In conversations, card numbers, passwords and one-time codes are hidden, and every time someone opens a conversation it is recorded.
What our team sees
Trucle's own operators work in a separate platform dashboard, with its own sign-in. It shows each store's plan, usage, invoices and connection health — the health as status codes and times, never your products, policies or conversations. Our team approves applications, manages plans and billing, and can pause or close a store. If your store's owner is locked out, our team can send a password reset, but it goes only to the owner's email address, we never see the password, and your dashboard tells you it happened.
Under the hood
- Tenant isolation in the database. Every table is behind a row-level policy with no exception path — not even for the application's owner role.
- The model cannot invent. It chooses prose, citations from the evidence it was given and products from the candidates it was given. Ids, prices, statuses and tool calls are the platform's.
- Measured behaviour. Intent extraction, injection resistance and grounded answering are scored against pinned datasets on the exact prompt and request production sends; a prompt edit that is not re-measured fails the build.
- Self-hosted model. Customer conversations are not sent to a third-party AI provider.
- Plans enforced at birth. A tenant is created with every tool outside its plan switched off, and its rate and concurrency limits follow its plan.