v0.0.3

v0.0.3 - Role-based knowledge loading for RAG

This release introduces native support for knowledge roles (system and user) in the loader and RAG engine pipeline.

Added

• role field in pkg/models.KnowledgeChunk.
• Role normalization during engine.ReadKnowledge:
  • system stays system
  • missing/invalid values default to user (backward compatible)
• Role propagation to all expanded chunks created from files and files_path.

Engine updates

• engine.Config now supports SystemPrompt string.
• RAGEngine now stores fixed system context and exposes it through SystemPrompt().

Reliability improvements

• ReadKnowledge now wraps underlying read/parse errors while preserving sentinel compatibility (errors.Is with ErrReadingKnowledgeFile and ErrParsingKnowledgeFile).
• Internal retrieval/filtering paths were tightened to avoid in-place slice mutation side effects.

Tests

• Added/updated tests for:
  • role propagation in expanded knowledge chunks
  • default role fallback to user
  • error assertions using errors.Is

Compatibility

• No breaking changes.
• Existing knowledge.json files without role continue to work unchanged.

Example

[
  {
    "id": "k1",
    "category": "persona",
    "files_path": "./kb/persona.md",
    "role": "system"
  },
  {
    "id": "k2",
    "category": "faq",
    "files_path": "./kb/faq",
    "role": "user"
  }
]

v0.0.2 - Hybrid Retrieval and Context Assembly

What's New

• Added hybrid retrieval mode combining dense and keyword search with Reciprocal Rank Fusion (RRF).
• Added hybrid + reranker mode (hybrid_reranker) for higher precision when reranking is available.
• Improved context generation with structured chunk headers (chunk_id, source, optional metadata), score filtering, and character-budget controls.

Improvements

• More robust context assembly with configurable limits:
  • ContextMax to cap total context size
  • ChunkMax to cap each chunk contribution
• Better retrieval defaults and candidate expansion logic for hybrid search quality.
• README updates and project naming consistency (gorag).

Fixes

• Reduced noisy embedding progress logging in ProgressEmbedder.
• Added safeguards to avoid empty/low-signal context output when scores are below threshold.

Tests

• Added/updated tests for hybrid retrieval behavior, retrieval mode normalization, and context character-budget enforcement.

v0.0.1 - Initial Release

Overview

v0.0.1 is the first stable public release of gorag, introducing a complete Retrieval-Augmented Generation (RAG) foundation in Go.

Highlights

• Introduced the core RAGEngine with retrieval and context-building flow.
• Added reranking support for improving final context relevance.
• Implemented document loading and chunk processing pipeline.
• Added concurrent progress-based embedding for batch operations.
• Added embedding and response caching layers for performance and cost optimization.
• Added Qdrant integration with collection bootstrap and management helpers.
• Added domain models for chat responses, knowledge chunks, and tool call logs.

Quality and Testing

• Added unit tests across core engine, cache, and store components.
• Included tests for cache embedder behavior and indexing correctness.

Developer Experience

• Added a working basic usage example under examples/basic.
• Improved documentation and project structure.
• Added MIT license and updated README content.