MultiTypeComponent is a base class that enables declarative, spec-driven component creation with automatic type registration, validation (required keys, type checks), and serialization (from_spec/to_spec). It is the foundation for all Lumen components (sources, views, pipelines, filters, transforms) and can be extracted as a standalone pattern for any Python project needing a plugin-like component system with YAML/JSON round-tripping.

A hierarchy of composable SQL transform classes (SQLLimit, SQLFilter, SQLPreFilter, SQLSample, SQLSchemaStats, etc.) that rewrite SQL queries using sqlglot with dialect-aware parsing and generation. Each transform implements a single responsibility (e.g., adding a LIMIT clause, injecting WHERE filters, sampling rows) and can be chained. The base class provides robust error handling, quoting, and dialect aliasing. This pattern is reusable for any project needing programmatic SQL manipulation with multi-dialect support.

FunctionTool wraps arbitrary Python functions as LLM-callable tools, automatically generating Pydantic schemas from function signatures and docstrings. It supports parameter sanitization, result accumulation across multiple calls, and integration with the LLM tool loop. This pattern is directly reusable for any LLM application needing to expose custom tools with minimal boilerplate.

An abstract VectorStore base class with a consistent API (add, query, upsert, delete, clear, filter_by) and three implementations: NumpyVectorStore (in-memory), DuckDBVectorStore (persistent with HNSW indexing), and ChromaDBVectorStore. Supports metadata filtering, chunking, and contextual augmentation ('situate'). The abstraction allows swapping backends without changing application logic, making it reusable for any RAG system.