Skip to main content

Compiled Wikis, Not Vector Embeddings

Lore builds persistent LLM knowledge bases from your project content — compiled markdown wikis organized by an LLM librarian. No vector embeddings. No retrieval noise. Just structured, human-readable, git-friendly knowledge that stays useful across sessions.

Built for teams who need their LLMs to retain real architectural context without the stateless reset of RAG.


Compiled Markdown Wikis, Not Vector Embeddings

Most knowledge base tools store your content as opaque vector embeddings — unreadable, uneditable, and locked to a specific retrieval model.

Lore compiles your content into structured markdown wikis. Read them. Edit them. Commit them to git. Your knowledge is yours, in a format that outlasts any model.

RAGLore
FormatVector embeddingsStructured Markdown
RetrievalSimilarity searchBacklinks + FTS5/BM25
PersistenceStatelessEvolving wiki + git
MaintenanceManualLLM-driven librarian

LLM-Driven Librarian

Lore doesn't just index your content — an LLM actively organizes and interlinks it. A 6-step compile pipeline extracts concepts, matches them to existing articles, generates line-level editing operations, and applies them with full provenance tracking. Every sentence knows where it came from.

New concepts are named, categorized, and cross-referenced. Orphaned pages are flagged. Ambiguities are surfaced. Multi-article edits and splits are handled automatically. It's like having a full-time research librarian maintaining your project's institutional knowledge.

lore init           # create a lore repo in your project
lore ingest ./docs # add source material
lore compile # LLM organizes and interlinks knowledge

Find exactly what you need without the noise of vector similarity search. Backlinks show you how concepts connect. FTS5/BM25 gives you fast, precise text retrieval. Query and search resolve through the graph and full-text index simultaneously.

lore search "architecture"
lore query "How does authentication work?"
lore path <concept> # show all paths to a concept

Code-Driven Pipeline

Lore's ingestion, compilation, indexing, and graph building are deterministic code. Your tokens go toward knowledge organization — understanding and linking concepts — not toward infrastructure overhead. No context windows burned on file I/O. No tokens wasted on serialization.

Incremental compilation skips unchanged content via manifest.json. A repository lock prevents overlapping runs. Every stage is optimized for token efficiency.


Paragraph-Level Provenance

Every article tracks exactly which sources contributed to which lines. Inline <!-- sources:HASH(CONFIDENCE) --> comments mark each paragraph's origin, while a cumulative ## References section records all sources ever merged. Provenance is organic — articles acquire it on first merge — with a --concepts-only flag available for backfilling existing wikis.

The auth service uses JWT tokens. <!-- sources:abc123(extracted) def456(inferred) -->

When the LLM reads articles for updates, provenance annotations are stripped so it sees clean, numbered text. The system manages provenance automatically — you just edit knowledge.


Mixed Source Ingestion

Lore normalizes content from everywhere your project knowledge lives:

  • Markdown, code files, and project docs
  • URLs and web pages
  • Chat transcripts (.json/.jsonl from supported agent frameworks)
  • Video transcripts (via yt-dlp)
lore ingest ./README.md
lore ingest https://example.com/architecture
lore ingest-sessions claude # ingest Claude Code session history

Export Everywhere

Your wiki isn't locked in a proprietary format. Export to whatever you need:

lore export --format slides
lore export --format pdf
lore export --format docx
lore export --format web
lore export --format canvas
lore export --format graphml

Presentations, documents, visual graphs — your knowledge goes where you need it.


Agent-Ready MCP Server

Lore ships with a first-class MCP server exposing 16 tools over stdio:

  • Retrieval: search, ask, explain, list_articles, get_article, get_neighbors, path
  • Graph diagnostics: graph_stats, lint_summary, list_orphans, list_gaps, list_ambiguous
  • Write: ingest, compile
  • Maintenance: check_duplicate, list_raw_tags, rebuild_index
lore mcp   # start MCP server for Claude Code, Cursor, VS Code Copilot, or any MCP host

Recommended agent loop: list_orphanslist_gapslist_ambiguousingest/compilerebuild_index(repair=true).


Git-Friendly & Portable

Your entire wiki is plain markdown files under .lore/wiki/. Commit it. Branch it. Include it in your project repo. Your knowledge travels with your code.

git add .lore/wiki/
git commit -m "Update project knowledge base"

Ready to Build Your Knowledge Base?

Get Started →

Or jump straight to Compiling Your Wiki and the CLI Reference.