How marianoorlei/qmd works
QMD is a local document-search engine exposed through a CLI, SDK, and MCP server. Users configure collections of Markdown files, index them, then retrieve documents or search by keywords, semantics, or hybrid ranking. The search path combines typed query expansion, BM25 and vector retrieval, reciprocal-rank fusion, and optional LLM reranking. Metadata filtering/discovery and model integrations are supporting capabilities. The repository also contains a separate model fine-tuning workflow; its internal steps are summarized at directory level because its sources were not inspected in full.
Main entry points: index.ts QMD SDK · store.ts Document indexing · qmd.ts QMD CLI · server.ts MCP server
Components 22
User interfaces
- QMD CLI
src/cli/qmd.ts - MCP server
src/mcp/server.ts - QMD SDK
src/index.ts
Collections and indexing
- Collection config
src/collections.ts - Document indexing
src/store.ts - Command processors
src/processors.ts
Search and ranking
- Query expansion
src/llm.ts - Hybrid search
src/store.ts - Fusion and reranking
src/store.ts - Local LLM models
src/llm.ts - Remote embeddings
src/remote-embed.ts - Remote reranker
src/jev.ts
Index and metadata
- Metadata extraction
src/metadata.ts - Metadata filters
src/metadata-filter.ts - Metadata discovery
src/metadata-store.ts - Document and vector store
src/store.ts - SQLite access
src/db.ts
Model fine-tuning
- Training data preparation
finetune/dataset - Reward modeling
finetune/reward.py - Model training
finetune/train.py - Evaluation
finetune/jobs
Other components
- User or agent
Connections 26
- User or agent → QMD CLI: runs commands
- User or agent → MCP server: connects tools
- User or agent → QMD SDK: calls API
src/index.ts - QMD CLI → Collection config: loads config
src/cli/qmd.ts - QMD CLI → Document indexing: reindexes files
src/cli/qmd.ts - QMD CLI → Command processors: creates processor
src/cli/qmd.ts - MCP server → QMD SDK: handles search
src/mcp/server.ts - MCP server → Collection config: loads config
src/mcp/server.ts - QMD SDK → Collection config: loads config
src/index.ts - QMD SDK → Document indexing: reindexes collection
src/index.ts - QMD SDK → Hybrid search: dispatches queries
src/index.ts - Document indexing → Document and vector store: writes index
src/store.ts - Document indexing → Metadata extraction: extracts metadata
src/store.ts - Document indexing → Command processors: runs processors
src/store.ts - Hybrid search → Query expansion: expands queries
src/store.ts - Hybrid search → Fusion and reranking: fuses results
src/store.ts - Hybrid search → Document and vector store: retrieves candidates
src/store.ts - Fusion and reranking → Local LLM models: reranks locally
src/store.ts - Fusion and reranking → Remote reranker: reranks remotely
- Query expansion → Remote embeddings: uses remote model
src/llm.ts - Metadata discovery → Metadata filters: compiles filters
src/metadata-store.ts - Metadata discovery → SQLite access: queries database
src/metadata-store.ts - Document and vector store → SQLite access: uses SQLite
src/store.ts - Training data preparation → Model training: prepares training data
- Reward modeling → Model training: scores training
finetune/train.py - Model training → Evaluation: evaluates models