Heta Framework v0.1.0

Heta Framework v0.1.0

Heta has completed its first framework-oriented release.

This version turns the original HetaDB knowledge-base workflow into a composable RAG framework. Developers can now build
different knowledge-base systems by assembling models, stores, parsers, steps, search modes, and benchmarks through a
KnowledgeRecipe.

Highlights

• Recipe-driven knowledge-base construction

  • Build a KnowledgeBase from explicit, reusable recipes.
  • Start with a simple vector KB, then add full-text search, graph construction, hybrid search, or benchmark evaluation
    as needed.

• Framework version of HetaDB-style graph building

  • Preserves the original HetaDB graph construction path.
  • Supports chunk merge, rechunk, entity extraction, relation extraction, deduplication, graph storage, and graph search
    as composable steps.

• Multiple search modes

  • vector_search
  • sql_text_search
  • full_text_search
  • heta_graph_search
  • hybrid_search
  • heta_rerank_search
  • heta_rewrite_search
  • heta_multihop_search

• Pluggable storage layer

  • Local object storage
  • S3-compatible object storage
  • In-memory and Milvus vector stores
  • SQLStore for SQLite/PostgreSQL/MySQL-style workflows
  • Elasticsearch-backed full-text index store

• Benchmark-oriented evaluation

  • BenchmarkRunner can build KBs from recipes, run query modes, and generate evaluation reports.
  • Current benchmark adapters include MultiHop-RAG, BEIR, and UDA-Benchmark related flows.

• Documentation site

  • Chinese and English docs are available through GitHub Pages.
  • Docs include Quick Start, Recipe, build paths, query modes, evaluation, stores, steps, and core components.

Links

• Documentation: https://knowledgexlab.github.io/Heta_Framework/
• Repository: https://github.com/KnowledgeXLab/Heta_Framework
• PyPI: https://pypi.org/project/heta-framework/

Notes

This is the first public framework release. The core build, search, graph, and evaluation paths are available in v0.1.0.
Future releases will continue to improve production recovery, error handling, benchmark coverage, and compatibility with
more RAG and GraphRAG methods.