Overview
The knowledge graph for the Latent Space community — two repos, one database, four surfaces.
The System
Two repos, one knowledge graph.
| Repo | What it is | Deployed on |
|---|---|---|
| latent-space-hub | Knowledge graph — Next.js web app, ingestion pipeline, MCP server, evals | Vercel |
| latent-space-bots | Discord bot (Slop) — agentic MCP tool-calling, member memory | Railway |
They share one Turso database. The hub writes to it (ingestion, web UI, API). The bot reads from it (MCP tool calls) and writes member data back.
How It Works
Content sources AI enrichment Surfaces
────────────────── ────────────────── ──────────────────
YouTube transcripts → Embeddings (1536d) → Web App (Next.js)
Substack articles → Chunk splitting → MCP Server (agents)
AINews (smol.ai) → Entity extraction → Discord Bot (Slop)
LatentSpaceTV → Auto-edge creation → Announcements webhook
- Content goes in — auto-ingestion pipeline polls RSS feeds hourly, extracts from YouTube and Substack
- AI enriches it — chunks are embedded, entities extracted, edges created automatically
- Humans and agents explore it — via the web UI, MCP tools, or Discord
- The graph grows — each new piece of content connects to existing knowledge
Architecture
Four cloud services, no self-hosted infrastructure.
| Service | Role |
|---|---|
| Vercel | Web app (read-only) + hourly ingestion cron + Discord announcements |
| Turso | Cloud SQLite — single shared database for web app, bot, and MCP |
| Railway | Slop bot — always-on Discord gateway with agentic MCP tool-calling |
| OpenRouter | LLM routing for Slop (currently Claude Sonnet 4.6, swappable) |
The Graph
| Metric | Count |
|---|---|
| Total nodes | ~3,900 |
| Content nodes | ~515 (podcasts, articles, AINews, workshops, paper clubs, builders clubs) |
| People & entity nodes | ~3,360 (guests, organizations, topics) |
| Edges | ~7,500 |
| Chunks with embeddings | ~35,800 |
| Coverage | June 2023 → present, continuously updated |
Tech Stack
| Layer | Technology |
|---|---|
| Framework | Next.js 15 + TypeScript + Tailwind CSS |
| Database | Turso (cloud SQLite via @libsql/client) |
| Search | Turso native vector search (F32_BLOB + vector_top_k) + FTS5 + hybrid RRF |
| Embeddings | OpenAI text-embedding-3-small (1536d) |
| AI | Anthropic Claude (entity extraction) + OpenAI (embeddings) |
| MCP | Model Context Protocol server — npx latent-space-hub-mcp |
| Bot | Discord.js + OpenRouter → Claude Sonnet 4.6 |
Built on RA-H, a local-first personal knowledge graph.