Integration

📘What it really is

• Put the parts together into one application, following the architecture from Module 81:

React assistant page (inside the CRM, behind login)
  → POST /v1/assistant/ask  (auth, role, rate and size limits, kill-switch flag)
      → standalone question rewrite (for follow-ups)
      → hybrid retrieval with permission and status filters → rerank → top passages
      → grounded prompt (versioned template) + read-only tools (customer summary)
      → model API (streaming, typed errors, cost logging)
      → checks: citations valid, numbers supported, faithfulness for flagged topics
      → stream answer + sources to the page; save to the conversation log (masked)
  → draft_followup_email tool → pending action → approvals page → send on approval
n8n: re-index on document change, review-date reminders, daily quality report

• Engineering standards from earlier levels still apply:
• TypeScript types and Zod validation throughout (Level 4);
• repositories and parameterised SQL (Level 5);
• tests: unit tests for chunking and checks; API tests for permissions and limits; the eval suite for quality (Module 89);

• configuration for model names, limits and prompt versions; secrets in the secrets store (Module 77).

• User experience:

• make it clear it's an AI assistant, with sources on every answer and a short note that answers can be wrong;
• suggested example questions; a "couldn't find it" answer that points to who to ask;
• feedback buttons; a way to report a harmful or wrong answer;

• fast feedback (streaming) and graceful messages when it's unavailable (Module 84).

• Accessibility and language: keyboard and screen-reader friendly (Module 37); test with the languages staff actually use.