minicloud Copilot

EPIC: #260 · Status: 📋 Backlog (5 connectors) · Sprint: TBD

minicloud Copilot is the self-hosted equivalent of Microsoft 365 Copilot. Every data source and AI component already exists on the platform — these 5 connectors wire them together so that AI assistance is embedded in the daily tools employees already use (email, chat, meetings, documents, search).

Stack mapping — minicloud vs M365 Copilot

M365 Copilot capability	minicloud equivalent	Status
Outlook — email summarization, triage	Stalwart mail + n8n + LiteLLM	📋 #261
Teams — meeting notes, action items	Jitsi + Whisper + LiteLLM + Nextcloud	📋 #262
SharePoint / OneDrive grounding	Nextcloud webhook → RAG → Qdrant	📋 #263
Copilot Chat in Teams	Matrix @ai bot → LiteLLM → Element	📋 #264
Enterprise Search	Unified RAG in Open WebUI (Qdrant backend)	📋 #265
AI gateway	LiteLLM	✅ live
AI observability	Langfuse	✅ live
Document conversion	markitdown-proxy	✅ live
Vector search	Qdrant	✅ live
Automation	n8n + minicloud-crew-agent	✅ live

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                        minicloud Copilot                            │
│                                                                     │
│  Data Sources              Connectors           AI Layer            │
│  ──────────                ──────────           ────────            │
│  Stalwart IMAP ──→ n8n ──────────────────────→ LiteLLM             │
│  Nextcloud files ──→ markitdown-proxy → RAG ──→   │                │
│  Jitsi recording ──→ Whisper ────────────────→    │                │
│  ERPNext API ────────────────────────────────→    │                │
│  Matrix history ─────────────────────────────→    ↓                │
│                                              Qdrant (vector DB)     │
│                                                   │                │
│                                              Open WebUI             │
│                                         (Enterprise Search)         │
│                                                   │                │
│                                              Langfuse (traces)      │
└─────────────────────────────────────────────────────────────────────┘

All connectors write to the same Qdrant instance. Open WebUI becomes the unified Enterprise Search interface — no separate search tool needed.

Connector 1 — Email AI (#261)

Stalwart IMAP → n8n → LiteLLM → Matrix + ERPNext + SMTP

Step	Component	Detail
Trigger	n8n IMAP node	Polls shared mailboxes (sinistres@, courtiers@, production@) every 2 min
Classification	LiteLLM	Urgency P1/P2/P3 + type (déclaration / résiliation / avenant / autre)
Extraction	LiteLLM	Insured name, policy number, date of loss, 3-sentence summary
Alert	Matrix webhook	Posted to `#sinistres:` room; P1 triggers @room
CRM creation	ERPNext REST API	Lead/Opportunity with extracted fields
Acknowledgement	SMTP via Stalwart	Auto-reply with SLA: P1=4h / P2=24h / P3=48h
Draft assist	LiteLLM	Reply draft saved to Stalwart Drafts (human sends)

Value for insurance IS: Eliminates manual triage of sinistres@ — every claim email is classified, logged in ERPNext, and acknowledged within 5 minutes.

Connector 2 — Meeting Notes (#262)

Jitsi → Whisper → LiteLLM → Nextcloud + Matrix

Step	Component	Detail
Recording	Jibri (to deploy)	Saves .mp4 to Nextcloud `Meetings/recordings/`
Trigger	n8n WebDAV poll	Detects new recording file
Transcription	Whisper large-v3	OpenAI-compatible endpoint in `ai` namespace
Summarization	LiteLLM	Structured output: participants, decisions, action items, summary
Storage	Nextcloud	Notes saved to `Meetings/notes/<date>-<room>.md`
Notification	Matrix	Summary posted to corresponding Matrix room

Action item format: - [ ] @owner: task description (due: YYYY-MM-DD)

Connector 3 — Document Grounding (#263)

Nextcloud webhook → markitdown-proxy → rag-ingest → Qdrant

This is the fastest connector to ship — no new containers, just wiring existing services.

Collection mapping:

Nextcloud folder	Qdrant collection	doc_type
/Company/Policies/	company-policies	policy
/Company/Procedures/	company-procedures	internal
/Company/Sinistres/	sinistres	claim
/Company/Contracts/	contracts	contract

Trigger: Nextcloud Flow app fires webhook on file create/update/delete → n8n → rag-ingest service.

Connector 4 — Chat AI Assistant (#264)

Matrix @ai bot → LiteLLM → Element Web

Deployed as maubot in chat namespace. Users mention @ai in any Element room.

Command	Action
`@ai <question>`	AI reply with last 10 messages as context
`@ai /search <query>`	RAG search across all Qdrant collections → top 3 sources
`@ai /summarize`	Summarize last 20 messages in the room
`@ai /action-items`	Extract action items from last 20 messages

For document questions (message contains ? + policy keywords) → RAG lookup first → LiteLLM with retrieved context.

Connector 5 — Enterprise Search UI (#265)

Open WebUI + Qdrant → unified search across all 4 data sources

Open WebUI is reconfigured to use the shared Qdrant instance (VECTOR_DB=qdrant) instead of its default Chroma. Five Knowledge Bases are pre-configured:

Knowledge Base	Collections	Access
📧 Emails	emails	management
📁 Company Documents	company-policies, procedures, contracts	all
🎥 Meeting Notes	meeting-transcripts	all
💬 Chat History	matrix-history	collab
🔍 All Sources	all collections (re-ranked)	management

Hybrid search: BM25 keyword + dense vector, French BM25 tokenizer already in the minicloud-open-webui custom image.

Source citation: every response includes filename + collection + chunk score.

Build order

Step	Issue	Effort	Depends on
1	Email AI (#261)	~3 days (n8n workflow + LiteLLM prompt)	nothing
2	Document grounding (#263)	~1 day (webhook + existing rag-ingest)	nothing
3	Chat AI (#264)	~2 days (maubot deploy + plugin)	nothing
4	Meeting notes (#262)	~4 days (Jibri + Whisper + n8n)	Jibri
5	Enterprise Search (#265)	~2 days (Open WebUI config + KB setup)	#261 #262 #263

Total estimate: ~12 days for full minicloud Copilot. No new infrastructure except Jibri (Jitsi recording) and Whisper (both lightweight).

Observability

Every LLM call in all 5 connectors flows through LiteLLM → traced in Langfuse with:

tags: connector name (email-ai, meeting-notes, etc.)
metadata: source document / email subject / room name
session_id: tied to the source event for full trace

A Grafana dashboard minicloud-copilot-v1 will show: calls per connector / p95 latency / classification distribution.

Stack mapping — minicloud vs M365 Copilot​

Architecture​

Connector 1 — Email AI (#261)​

Connector 2 — Meeting Notes (#262)​

Connector 3 — Document Grounding (#263)​

Connector 4 — Chat AI Assistant (#264)​

Connector 5 — Enterprise Search UI (#265)​

Build order​

Observability​