Process for extracting packages/graph from the research monorepo and publishing Lattice as an opensource npm package.

Key Value Propositions

  1. Uses Claude Code subscription - Entity extraction runs through Claude Code, so users don’t pay for a separate LLM API
  2. Two-command workflow - /research finds or creates docs, /graph-sync extracts entities and syncs automatically
  3. Zero external dependencies - Embedded DuckDB, one API key (Voyage AI for embeddings only), 2 minutes to start

The Workflow

Terminal window
/research "topic"              # Find existing docs or create new research
/graph-sync                    # Extract entities & sync to database
lattice search "query"         # Semantic search your knowledge base

Documents

DocumentDescription
ArchitectureTechnical deep-dive: NestJS structure, DuckDB schema, sync workflow, Claude Code integration
CLI Search CommandsTechnical comparison of lattice CLI search commands: semantic search, SQL queries, relationship lookup
Frontmatter-Free ProposalHistorical proposal to refactor Lattice to eliminate YAML frontmatter (now implemented)
LightRAG ComparisonComparison with LightRAG: databases, UX, architecture
Repository InitializationGuide for setting up the Lattice repository

Architecture Summary

  • Database: Embedded DuckDB (zero Docker, zero Redis)
  • Vector Search: DuckDB VSS extension (HNSW index, cosine similarity)
  • Embeddings: Voyage AI (voyage-3-lite, 512 dimensions)
  • Runtime: Bun + NestJS + nest-commander
  • No frontmatter: Entities extracted directly to database, markdown stays clean