Skip to main content
Enterprise onboarding onto MITHUNAI follows a structured six-phase journey that transitions your organization from initial workspace provisioning to production multi-channel deployment with verified factual grounding. Each phase establishes explicit boundaries, security controls, and verification gates before moving to the next.

​
Phase 1: Workspace Setup & Member Provisioning

Every enterprise engagement begins with an isolated tenant organization governed by role-based access control (RBAC).
1

Verify Organization Provisioning

Your deployment administrator provisions your organization and assigns the initial Owner. Owners hold full administrative authority over members, collections, assistants, API keys, and branding.
2

Invite Team Members & Set Roles

Invite your core project team via the console or the members API:
  • Administrators: Knowledge engineers and security leads who configure connectors and manage API keys.
  • Editors: Technical writers and product managers who curate assistants and evaluate answers.
  • Members: Team members who query the assistant and review conversation histories.
Add Member via API

​
Phase 2: Connect Knowledge Sources

Connect the repositories, portals, and documentation spaces that form your factual ground truth.
  • Developer Documentation & Sitemaps: Enter your public or internal sitemap URL (https://docs.example.com/sitemap.xml). MITHUNAI validates the safe-fetch boundary and establishes crawl limits.
  • GitHub & GitLab Repositories: Connect your engineering repositories, specify branch targets (main or release/v2), and apply path globs (docs/**, api/**/*.ts).
  • Workspace Wikis (Notion & Confluence): Authorize enterprise OAuth to sync engineering spaces while preserving document hierarchies.
  • Document Archives: Drag-and-drop architectural whitepapers, PDF runbooks, and OpenAPI specifications.

​
Phase 3: Ingestion, Chunking & Verification

Trigger an asynchronous sync job to ingest, redact secrets, and vectorize the content into your organization’s partitioned pgvector store.
  1. Secret & Credential Redaction: Automated pre-index filters detect and mask API tokens, private SSH keys, and connection strings.
  2. Structural AST & Semantic Chunking: Markdown and code are split along logical boundaries (headers, classes, functions) rather than arbitrary byte boundaries.
  3. Ingestion Verification: Monitor job progress via GET /knowledge/jobs/{id} or the console dashboard. Ensure all pages commit with status completed.

​
Phase 4: Configure Assistant & Inference Gateway

Create and tailor your AI assistant for your specific audience.
  • Model Selection: Choose your preferred upstream model or configure private on-premise inference (vLLM, Ollama).
  • Prompt Fencing: Enable cryptographic nonces to isolate retrieved passages from system commands.
  • Semantic Caching: Activate the sub-45ms cache to serve repeated technical questions instantly at zero inference cost.

​
Phase 5: Grounding & Truth Evaluation

Before public deployment, validate answer quality against your domain’s benchmark questions in the Ask & Test Playground:
  1. Direct Evidentiary Checks: Verify that answers cite the exact line numbers and URLs of the source documentation.
  2. Abstention Verification: Test out-of-scope questions (e.g., “How do I configure product X that we don’t build?”). Ensure the assistant explicitly declines rather than hallucinating.
  3. Citation Click-Through: Click each generated citation badge in the test UI to confirm the highlighted excerpt matches the assertion.

​
Phase 6: Multi-Channel Production Launch

Deploy your assistant across the channels where your users and developers interact:

Embeddable Web Widget

Generate an arukz_wk_ key, configure your website’s origin allowlist (https://example.com), and paste the <script> tag onto your documentation portal.

REST API Integration

Mint an arukz_sk_ service key with rate limits to build conversational intelligence directly into internal portals or ticketing bots.

Model Context Protocol

Distribute the read-only MCP configuration to your engineering team for integration into Cursor, Claude Desktop, and Windsurf IDEs.
Last modified on September 27, 2026