Heta Framework v0.1.3b1

Added

• Added the optional generation package with a typed task-producer protocol,
  explicit producer configuration and input requirements, and object-store task
  manifests and reports.
• Added HetaGenTaskProducerExploration for source-grounded task authoring,
  building, review, and bounded revision.
• Added HetaGenTaskProducerDerivation for generating task variants from seed
  tasks through NeMo Data Designer.
• Added HetaGenTaskProducerGraph for evidence-grounded task generation through
  the native Omega workflow, accepting either a Heta knowledge base and query
  or prepared text and knowledge-graph objects.
• Added Harbor task packaging, static checks, and runtime validation with
  Oracle and Nop agents, together with bilingual usage and extension guides.

Changed

• Updated LiteLLM from 1.87.1 to 1.101.0.
• Added pinned Data Designer 0.9.2 and Harbor 0.23.0 dependencies to the
  generation extra; base installations do not install these dependencies.

Fixed

• Preserved original-document provenance in aggregated LightRAG query results.
• Normalized Graph producer stage summaries before runtime processing so mixed
  Parquet schemas do not discard records.
• Preserved explicit failure results across runtime stage boundaries and
  reported missing stage records alongside independently validated tasks.

Compatibility

• Base Heta Framework supports Python 3.10 through 3.12. Task generation requires
  Python 3.12 and the generation extra.
• Existing knowledge-base build and query APIs remain available.
• Runtime task validation requires Docker. Exploration additionally requires a
  locally prepared role image; installing the Python extra does not create it.

Known Limitations

• Task generation is a Beta API and generated content requires review for
  scientific correctness and task difficulty.
• Candidates that fail validation are excluded; a run may produce no tasks.
• Graph runtime validation failures do not trigger a complete automatic
  authoring and revalidation loop.

Installation

pip install "heta-framework[generation]==0.1.3b1"

Task generation requires Python 3.12 and Docker.

Full changelog: v0.1.2...v0.1.3b1

Heta Framework v0.1.2 — Unified Graph RAG Procedures

Heta Framework v0.1.2 — Unified Graph RAG Procedures

Release notes:

Heta Framework v0.1.2 introduces a unified graph extraction architecture and built-in support for multiple Graph RAG
approaches.

Highlights

Unified Graph Extraction

Graph construction now starts from a shared universal graph representation. GraphRAG, LightRAG, HiRAG, LeanRAG, and Heta
Graph can reuse the same extraction results and adapt them to their own graph structures.

The previous entity-then-relation extraction path has been replaced by:

• Universal graph extraction
• Optional ontology constraints
• Method-specific graph adapters
• Independent graph building and indexing steps

New Graph RAG Procedures

This release adds built-in procedures and query engines for:

• GraphRAG
• LightRAG
• HiRAG
• LeanRAG

Each procedure integrates its corresponding graph construction, storage, indexing, and retrieval workflow into the Heta
Framework Recipe system.

Hierarchical Graph Capabilities

New graph capabilities include:

• Leiden graph clustering
• Clusterable graph-store protocols
• Hierarchical graph aggregation
• Community report generation
• LeanRAG semantic aggregation
• In-memory graph clustering

Self-Contained LightRAG and LeanRAG Support

LightRAG and LeanRAG prompts are now bundled directly with Heta Framework.

Installed packages no longer depend on adjacent LightRAG or LeanRAG source repositories, enabling clean standalone
installation and deployment.

Multi-Document Provenance

Retrieval results can now be associated with multiple source documents. BEIR evaluation handling has also been updated to
support multi-document IDs correctly.

This release also fixes MultiHop-RAG evidence recall evaluation for aggregated HiRAG results. EvidenceRecallAtK now
supports both singular and plural provenance fields:

• source_key / source_keys
• object_key / object_keys
• document_id / document_ids

This enables aggregated results from GraphRAG, HiRAG, and LeanRAG to be evaluated correctly.

Installation

Install the base framework:

pip install -U heta-framework==0.1.2

Install graph construction and Graph RAG dependencies:

pip install -U "heta-framework[graph]==0.1.2"

Compatibility

• Supports Python 3.10, 3.11, and 3.12
• Existing Wiki and agentic query APIs from v0.1.1 remain available
• Includes compatible versions of NetworkX, graspologic, NumPy, scikit-learn, and UMAP
• Python 3.13 is not supported in this release

Full Changelog:
v0.1.1...v0.1.2

Heta Framework v0.1.1

Heta Framework v0.1.1

This release introduces Heta Wiki construction and agentic retrieval while preserving compatibility with existing v0.1.0
knowledge bases and recipes.

Highlights

• Added ToolCallingLanguageModelProtocol and the LiteLLM-backed ToolCallingLanguageModel.
• Added BuildWikiPages, SplitWikiPages, and HetaWikiProcedure.
• Added Wiki presets for chunk embedding, vector indexing, and full-text indexing.
• Added reusable Wiki query tools:
  • read_wiki_index
  • search_wiki
  • read_wiki_page
  • read_raw_object
• Added AgenticQueryEngine and the wiki_agent_search query mode.
• Added evidence-backed QueryInsight output.
• Preserved Markdown-oriented parser content through optional original_content.
• Propagated Wiki page and original-document provenance through search results, citations, and retrieval evaluation.

Heta Wiki Pipeline

ParseDocuments
→ BuildWikiPages
→ SplitWikiPages
→ EmbedChunks(preset="wiki")
→ IndexVectors(preset="wiki")
→ IndexFullText(preset="wiki")

The resulting knowledge base supports:

- wiki_vector_search
- wiki_full_text_search
- wiki_agent_search

### Compatibility

- Existing v0.1.0 parsed documents and chunks remain readable.
- Existing knowledge bases do not require data migration.
- Existing positional configuration constructors retain their v0.1.0 field order.
- Existing recipes and default query modes remain unchanged.
- Wiki construction and agentic retrieval are opt-in.
- QueryResponse.insights is appended without changing existing field positions.

### Behavior Change

The default EmbedChunks batch size is reduced from 64 to 10, improving compatibility with embedding providers that enforce
smaller batch limits. Users may explicitly increase it when supported by their provider.

### Validation

- 274 tests passed, 11 skipped
- Ruff passed for all files changed since v0.1.0
- MkDocs strict build passed
- Wheel and source distribution built successfully
- twine check passed
- Isolated wheel installation and public API smoke tests passed
- Real PDF and BEIR SciFact Wiki pipelines were validated

### Installation

pip install --upgrade heta-framework==0.1.1

Optional integrations:

pip install "heta-framework[milvus,s3,text-index,postgres]==0.1.1"

### Known Limitations

- Adding documents to an already successful Wiki knowledge base remains an explicit update lifecycle.
- Resuming a build requires a recipe equivalent to the one used for the original run.

Full Changelog: https://github.com/KnowledgeXLab/Heta_Framework/compare/v0.1.0...v0.1.1

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.