Generated from the build.

Interfaces

• Microsoft Teams / Azure Bot Service: inbound staff messages arrive as Bot Framework activities at the n8n webhook; answers are sent back to {serviceUrl}/v3/conversations/{id}/activities.
• n8n: two workflows, ops-assistant.query.json (router, specialists, reply) and ops-assistant.ingest.json (per-domain ingestion).
• Anthropic Claude API: reasoning model for the router and specialists, via the Anthropic Chat Model node.
• OpenAI API: embeddings (text-embedding-3-small) for ingestion and query.

Authentication

• Teams: an Azure Bot registration; outbound replies use a Bot Framework bearer token (Azure AD client credentials). Least-privilege Microsoft Graph scopes.
• Claude and OpenAI: API keys held in n8n credentials, never in the workflow JSON.

Retrieval contract

router:       classify message, delegate to exactly one specialist
specialists:  answer only from their own vector store; decline if not documented
ingestion:    SOP files -> recursive split (512 / 50 overlap) -> embed -> per-domain store
query:        specialist retrieves from its store as a tool (retrieve-as-tool, topK 4)
model:        one shared Claude model + one shared embeddings node across all agents

Resilience

• The webhook returns 200 immediately so Teams never hits its activity-response timeout; the answer follows asynchronously.
• Knowledge isolation per domain prevents cross-domain answers.
• Undocumented questions are declined rather than guessed.
• Documents are retrieved at query time and never used to train or fine-tune a model.

Setup and reconfiguration

1. Run n8n; add Anthropic and OpenAI credentials.
2. Import both workflows.
3. Place the SOP files where the ingestion Read nodes point, then run ingestion once.