How marker-inc-korea/autorag works
AutoRAG Agent is a self-evolving librarian for searching documents in place, verifying evidence, and returning curated findings; the repository also retains the legacy Python AutoRAG tool, which is not the main product mapped here. The documented workflow consults memory, searches through multiple retrieval methods, applies access controls and merges results, then directly reads promising sources and emits findings. The graph preserves the documented MinSync, Jikji, datasource, web, Lite, and feedback capabilities. Several subsystem links follow README descriptions because their implementation was not sampled.
Main entry points: agent.ts Librarian Agent · engine.ts Retrieval Engine · factory.ts Datasource Skills · index.ts CLI
Components 20
User Entry
- CLI
src/cli/index.ts - Librarian Agent
src/agent/agent.ts - Evolving Memory
src/memory/memory.ts
Retrieval
- Retrieval Engine
src/retrieval/engine.ts - Method Registry
src/retrieval/registry.ts - MinSync Search
src/minsync/client.ts - Jikji Discovery
src/jikji/client.ts
Data Sources
- Datasource Skills
src/datasource/skills/factory.ts - Crawler Client
src/datasource/crawler-client.ts - Document Mirror
src/mirror/sync.ts - Document Parser
src/parser/index.ts
Local Runtime
- Embedding Gateway
src/embedding-runtime/index.ts - Lite Commands
src/cli/commands/lite.ts
Research & Feedback
- Direct Evidence Read
src/agent/bash-tool.ts - Web Search & Fetch
src/agent/web-search-tool.ts - Feedback
src/cli/commands/feedback.ts - Curated Findings
src/agent/emit-results-tool.ts
Other components
- User
- Local Documents
- External Data Sources
Connections 22
- User → Librarian Agent: asks
- User → CLI: runs command
- CLI → Lite Commands: dispatches
- Librarian Agent → Evolving Memory: consults
- Librarian Agent → Retrieval Engine: searches
- Retrieval Engine → Method Registry: selects methods
- Retrieval Engine → MinSync Search: queries
- Retrieval Engine → Datasource Skills: queries
- Librarian Agent → Jikji Discovery: discovers
- Datasource Skills → Crawler Client: invokes
- Crawler Client → External Data Sources: searches
- External Data Sources → Datasource Skills: returns hits
- MinSync Search → Embedding Gateway: uses embeddings
- Document Mirror → Document Parser: parses documents
- Local Documents → Document Mirror: syncs documents
- Librarian Agent → Direct Evidence Read: reads evidence
- Direct Evidence Read → Local Documents: opens sources
- Librarian Agent → Web Search & Fetch: researches web
- Librarian Agent → Curated Findings: emits findings
- User → Feedback: provides feedback
- Feedback → Evolving Memory: records outcomes
- Lite Commands → Retrieval Engine: retrieves