Jorvis North Star

Version: v0.8.7 Owner: George (Release Authority) Updated: 2026-03-28

Mission

Jorvis — chat with any data. A local-first platform that lets operators and analysts explore any database through natural language, without sending data to third-party AI services.

Primary Users

User	Primary Need
Business operator	Ask questions about their data without writing SQL
Data analyst	Accelerate exploration of unfamiliar schemas
Investor / demo audience	See AI-driven data insight, live, on real data

Current Demo / Pilot Target (next 30-60 days)

Primary demo target: AdventureWorks dataset plus the restored Cars/SQL demo surface on jorvis.app Primary journey: Natural language query → SQL → answer + Transparency panel Secondary journey: Document upload → OCR → Knowledge Base Q&A

In Scope (now)

Natural language → SQL → answer (Text2SQL)
Graph-augmented SQL (GraphRAG — plain-SQL, graph_edges)
Document RAG (OCR + Knowledge Base)
Provider-backed Open WebUI text models
OpenClaw capability routing and memory-search
Local LLM (jorvis-4b-256k, offline, GPU)
Cloud LLM (Gemini 2.5+)
Voice (STT/TTS)
Open WebUI integration

Explicitly Out of Scope (now)

Multi-tenant SaaS
User authentication / access control beyond admin
External database connectors beyond configured JDBC/psql
Mobile-native app
Real-time streaming from external data sources
General image generation chat-row surface; the supported future path is native Open WebUI Images backed by an OpenAI-compatible image backend

"What Good Looks Like" (next 30-60 days)

Investor demo runs end-to-end without errors on prod (AdventureWorks plus the restored SQL demo surface)
GraphRAG: graph_edges synced, expansion visible in Transparency panel
OCR: PDF upload → successful KB ingestion → correct Q&A response
Voice: STT → query → TTS response working on jorvis.app
CI: quality gate green, self-hosted runner stable, coverage non-regression baseline intact
Help KB: complete across Jorvis docs, Open WebUI docs, and OpenClaw docs

Exploratory Next Slice (Not Yet GO)

Candidate: Gemini Embedding 2 multimodal pilot for document/image retrieval
Intent: Extend the secondary journey from text-only OCR/KB behavior toward provenance-rich multimodal retrieval
Bounded shape: PDF + page images + OCR fallback + image-aware retrieval, behind feature flags and separate storage
Non-goal: No full RAG rewrite, no external vector DB migration, no broad video lane in the first slice
Planning docs:
- docs/product/GEMINI_EMBEDDING_2_JORVIS_FIT_ANALYSIS.md
- docs/product/GEMINI_EMBEDDING_2_MULTIMODAL_PILOT_EXECUTION_PLAN.md
- docs/agent_ops/specs/task_gemini_embedding_2_multimodal_pilot_stage0_spec.md

What This Document Is NOT

Not a task board → see docs/agent_ops/TASK_BOARD.md
Not a release log → see docs/handoff/HANDOFF_TO_NEXT_AGENT.md
Not an architecture doc → see docs/architecture/
Not a roadmap with dates → see docs/handoff/ROADMAP_v0.8.0.md