What happens to your message
AllFrom your point of view the chat is simple: question in, answer out. Behind the scenes, though, every message goes through several stages — it is stripped of personal data before sending, the knowledge base is searched, the assistant calls tools when needed, and only then does the answer take shape. This page shows the complete journey in order and — as a common thread — where you can follow each stage in the Activity sidebar.
The five stages of a message
Section titled “The five stages of a message”1. PII protection (the PII shield)
Section titled “1. PII protection (the PII shield)”Before anything goes to the AI model, CompanyWizard scans your message for personal data (names, email addresses, phone numbers, postal addresses …) and replaces each match with a placeholder, e.g. Max Mustermann → [PERSON_1]. Only this pseudonymized text leaves your instance. When the answer comes back, the placeholders are automatically swapped back to the original values.
- Your own message stays in the history as the original — only the copy the model sees is pseudonymized.
- File attachments too (text and PDF) are scanned before processing.
- If the shield is deliberately bypassed for one case (“bypass PII”), that appears as its own entry in the activity.
Go deeper: The PII shield.
2. Knowledge search (RAG)
Section titled “2. Knowledge search (RAG)”Before answering, CompanyWizard searches your knowledge base and passes the matching passages to the model as context — so the AI answers based on your documents, not just general knowledge. The matches are filtered by your access rights: you only get context from content you are allowed to see anyway.
If no semantic vector search is currently active, a classic keyword search runs instead — recognizable by the “Local index” hint in the Activity sidebar. Match quality is lower in that case.
Go deeper: RAG & search · RBAC & confidentiality levels · Knowledge base & vaults.
3. Tools
Section titled “3. Tools”While working, the model can call tools — for example the knowledge search, web search, the workspace folder, data sources configured by administrators, skills, or memory. Tools run one after another: each call appears live in the activity (first “running”, then “done”) before the next starts. That lets you see why an answer takes a little longer, or where a piece of information came from.
Go deeper: The assistant’s tools — overview · Skills, memory & follow-up questions.
4. Answer streaming
Section titled “4. Answer streaming”The answer appears token by token as it is generated — you don’t have to wait for the full text. As it streams, the placeholders set in stage 1 are swapped back to the original values, so you read the correct names and details throughout.
5. A follow-up question (the assistant asks back)
Section titled “5. A follow-up question (the assistant asks back)”Sometimes the model is missing a detail it needs to continue sensibly. In that case it can ask a question mid-process. The chat pauses and waits for your answer — then continues exactly where it left off, without starting over. Context already gathered (knowledge matches, tool results) is preserved.
The Activity sidebar — your window into the process
Section titled “The Activity sidebar — your window into the process”The “Activity” tab shows the stages as a chronological list, top to bottom in the order they happen. As soon as an answer begins, the view switches to this tab automatically so you can follow along right away.
Each entry carries a colored dot:
| Dot | Meaning |
|---|---|
| 🟢 Green | Knowledge search succeeded — matching results found |
| 🔴 Red | Search failed or returned nothing |
| 🟡 Amber, pulsing | A tool is currently running |
| 🔵 Blue | Tool finished (with duration) |
Other elements:
- “Local index” chip — appears when the simple keyword search is active (no vector search).
- Clear — resets the list for the next process.
- On mobile, the same sequence appears as an expandable inline dropdown right in the chat instead of a sidebar.
A worked example
Section titled “A worked example”You ask: “Summarize the contract with Max Mustermann and check whether the current termination date is correct.”
In the “Activity” tab you would see, in order, something like:
- 🛡️ PII shield — 1 match pseudonymized (
Max Mustermann→[PERSON_1]); only the placeholder goes to the model. - 🟢 Knowledge search — pre-search in the knowledge base, matching contract passages found (rights-filtered).
- 🟡 Tool running — the assistant opens the contract document …
- 🔵 Tool done — document read (e.g. 0.4 s).
- ✍️ Answer — the summary appears token by token;
[PERSON_1]turns back into Max Mustermann on display.
If the model had needed a missing detail (say, which of two contracts you meant), a follow-up question would have appeared after step 2 — the process would have paused and resumed at step 3 after your answer.
What the model sees — and what it doesn’t
Section titled “What the model sees — and what it doesn’t”- The AI model only ever receives the pseudonymized text plus the rights-filtered knowledge context. It never sees real names or contact details.
- Your chat history shows the original values throughout — your typed message and the re-identified answer.
- The activity list records which stages ran, not the confidential content itself.
That is the heart of the trust promise: you work with real data, the model works with placeholders.