Agent-first academic research knowledge service — a knowledge flywheel for CS/AI researchers.

NORIA transforms scattered literature into structured, provenance-tracked, citation-linked knowledge. It auto-discovers papers, compiles them into an interlinked wiki, detects knowledge gaps, and serves the results via MCP to external agents.

• 5-level provenance model — every claim carries trust level (user-verified > source-derived > llm-derived > social-lead > query-derived). Synthesis requires ≥2 source-derived citations.
• Knowledge flywheel — feedback → gap detection → source expansion → better retrieval → feedback. Closed-loop via signal-index + demand prior reranking.
• Section-level citations — [source: citekey, sec.3.2], not just paper-level references. Venue claims verified against S2/DBLP/OpenReview.
• 23 slash commands — full CLI workflow from /kb-sync (discovery) to /kb-reflect (synthesis) to /kb-expand (flywheel automation).
• 30 TypeScript/Python tools — progressive PDF reader, multi-platform search (arXiv, S2, GitHub, Twitter, WeChat), graph analysis, community detection.
• MCP Knowledge Service — remote agents query the wiki via search, ask, gap_scan, list_concepts, graph_neighbors, submit_feedback.
• Multi-model routing — Opus (synthesis), Sonnet (compile), Haiku (lint), GPT-5.4 (adversarial review).
• Obsidian frontend — Juggl graph visualization (4-layer node coloring), Dataview dashboards, MOC pages, Canvas maps.

Raw Sources (user owns)  →  LLM Engine (Claude Code)  →  Wiki (LLM maintains)
   Zotero / arXiv / S2        23 skills + 30 tools         sources / concepts /
   Twitter / WeChat / GitHub   multi-model routing          synthesis / entities
                                                         →  MCP Service (remote agents)

See ARCHITECTURE.md for the full system design, directory structure, provenance model, and tool inventory.

• Claude Code CLI
• Node.js 20+ with npx tsx
• Python 3.10+ (for Zotero sync, MCP server)
• Obsidian (optional, for visualization)

git clone https://github.com/fzhiy/noria.git
cd noria

# Start working with Claude Code
claude

# Show project overview
/wiki-help

# Sync latest papers from Semantic Scholar
/kb-sync s2 "web agent reinforcement learning" --limit 10

# Compile new sources into wiki
/kb-compile

# Check wiki health
/kb-lint

# Ask a research question
/kb-ask "How does WebRL handle curriculum generation?"

# Detect knowledge gaps
/kb-gap-scan

# Write cross-cutting synthesis
/kb-reflect

# Start the MCP server
python3 tools/noria-mcp-server.py 3849

# From another project, configure in .mcp.json:
# { "noria": { "url": "http://localhost:3849" } }

1. Provenance-first — every claim traceable to source with trust level
2. Token-efficient — progressive reading (5 modes), manifest-gated compile, RT pre-screening
3. Human-gated — auto-discovery but human-approved expansion
4. Lint before reflect — deterministic quality gate before LLM synthesis
5. Multi-model routing — cheapest sufficient model for each task
6. Dual-track isolation — social media (social-lead) quarantined from academic synthesis

• ARCHITECTURE.md — full system design, tools inventory, directory structure
• schema.md — wiki page format, provenance rules, frontmatter specification
• docs/tooling-reference.md — tool usage details
• docs/juggl-visual-guide.md — Obsidian graph visualization guide
• docs/remote-wiki-access.md — MCP remote service setup

Extends the Karpathy llm-wiki pattern with provenance tracking, multi-model adversarial review, progressive PDF reading, and dual-track information architecture.

MIT