The product
A conversational assistant embedded in your web application. Three modes, one navigation model, continuous checks.
The navigation model
NAVAGE walks through your application and builds its model: screens, available actions, flows, content. You review and adjust it with us.
This model is the agent’s only source of truth. It contains only what exists.
Guide
The agent highlights the element to activate, step by step, or performs the action on the user’s behalf when you allow it. The user stays on their page, in your interface.
Take
“I want to change my address”: the agent computes the route from the model and takes the user straight to the right screen. Nothing is guessed from what is displayed.
Answer
Questions are answered only from your site’s content, with the source shown. Without a reliable passage, the agent says so: “I don’t have that information here”. Unanswered questions are reported to you so you can enrich your content.
The language model: how, why, at what cost
How: every request first goes through the free stages, recognised commands then local matching against the navigation model. The language model is called only as a last resort, to understand an unexpected wording or to write an answer from the passages of your site.
Why: because human language is varied, and that is the only thing a model does better than a rule. Everything else, the route to follow, the element to activate, the refusal of a non-existent flow, is computed, not generated.
At what cost: on our test sets, 79 to 87% of requests are resolved without calling the model. When a call happens, it is short and single. A session costs next to nothing; we do not bill it. NAVAGE is provider agnostic: yours, ours, or none.
Continuous checks
On every release, the model is verified against the real application. A modified screen, a missing action: your team is warned before your users.
Completion metrics
Completion rate per flow, drop-off per step, unresolved requests. You see where users get stuck, with no extra analytics tool.
Languages and voice
French and English. The session language follows your application’s, declared, never guessed. One female voice per language; read-aloud and dictation can be enabled by the user.
Integration
One line added to your site, no change to its code, session kept from page to page. Details on the Integration page.
What we measure
- On our internal evaluation sets (24 questions, 2 sites), no invented answer: the assistant answers from the content or says it does not know.
- 79 to 87% of requests resolved without calling a language model on our test sets.
- About 0.9 s per answer with a language model, on our test sets.
Roadmap
- Mobile applications: planned.
- Spanish and Arabic: planned.
- Client console: in development.
What the product does not do
- It does not read your users’ screen.
- It does not offer an action missing from the model.
- It does not answer without a source.
- It does not collect more than the request, the language and the flow state.