Skip to content

Release Notes

Stay up to date with the latest features, improvements, and bug fixes in Maticlib.

[v0.2.2] - 2026-05-03

Major Additions

Formatting & Evaluation

  • Markdown Results: Convert SQL tuples and columns into a clean Markdown table format.
  • Evaluation System: Flexible test modules (ContextRelevanceEvaluator, AnswerAccuracyEvaluator) to grade outputs with fallback heuristics and LLM options.

[v0.2.1] - 2026-05-03

Major Additions

Prompts, Observability & Resilience

  • Prompts & Memory: Established BasePromptTemplate and dynamic PromptRegistry coupled with rolling window ConversationBufferMemory.
  • Observability & Resilience: Built high-resolution tracing (PipelineTrace, StepTrace) natively capturing token budgets and a robust exponential backoff @with_retry decorator.

Vector Stores & Pipelines

  • Vector Stores: Built integrations for ChromaVectorIndex, MilvusVectorIndex, PineconeVectorIndex, QdrantVectorIndex, and InMemoryVectorIndex.
  • Advanced Retrieval: Implemented HybridRetriever, ReciprocalRankFusion (RRF), and LLM-backed QueryTransformer.
  • Pipelines: Unified the logic into easy-to-use RAGPipeline and Text2SQLPipeline orchestrators.

[v0.2.0] - 2026-05-03

Major Additions

RAG & Context Engineering (Phase 2)

Completed a massive expansion of the Maticlib data pipeline, enabling advanced RAG and Text2SQL workflows.

  • Advanced Chunkers: Built SeparatorChunker, HierarchicalChunker, TokenBudgetChunker, and SemanticDifferenceChunker for precise context engineering.
  • Document Loaders: Added robust TextLoader, PDFLoader, DOCXLoader, and WebPageLoader to ingest data seamlessly into the chunking system.
  • Vector Stores Config: Implemented VectorIndexConfig defining Flat, HNSW, IVF, and LSH indexing strategies.
  • Text2SQL Engine: Introduced full schema reflection, dialect transpilation, tabular file ingestion (CSV/Excel/Parquet), and the powerful SQLInjectionGuard utilizing sqlglot to secure LLM-generated queries.

[v0.1.9] - 2026-05-01

Major Additions

Embeddings Token Tracking & Metadata

  • Changed embed_query() and embed_documents() return types to return Pydantic objects (EmbedQueryResponse and EmbedDocumentsResponse).
  • Includes token usage (prompt_tokens, total_tokens), precise model IDs, and the raw provider response dictionary directly inside the returned model.

[v0.1.8] - 2026-04-29

Major Additions

RAG & Context Engineering (Phase 1)

Introduced a unified interface for generating high-dimensional vector representations of text.

  • Unified Interface: New BaseEmbeddings abstract class ensuring a consistent embed_query and embed_documents API across providers.
  • OpenAI Embeddings: OpenAIEmbeddings client supporting text-embedding-3 models with dimensionality controls.
  • Google GenAI Embeddings: GoogleGenAIEmbeddings client supporting gemini-embedding-001 with task-type optimization.
  • Mistral Embeddings: MistralEmbeddings client supporting the mistral-embed model.

[v0.1.7] - 2026-04-21

Major Additions

Native Tool Calling (Function Calling)

A robust system for enabling LLMs to interact with external Python functions. Developers can now define tools using a simple decorator and pass them directly to any LLM client.

  • @tool Decorator: Automatically generates JSON Schemas from Python function signatures, docstrings, and type hints.
  • Multi-Provider Support: Standardized implementation across OpenAI, Google Gemini, and Mistral AI.
  • Unified Response Interface: Tool requests are automatically extracted into a common tool_calls field in the standardized response model.

Fixes & Improvements

  • Google GenAI: Updated default model to gemini-2.5-flash-lite for better stability and lower latency.
  • OpenAI: Restored missing response_id and model_version mappings in the Responses API.
  • Imports: Fixed a missing Callable import in several LLM client modules to prevent runtime errors during test collection.

[v0.1.6] - 2026-04-17

Major Additions

Response Parser System

A new, LangChain-inspired structured output system. Developers can now pass a Pydantic BaseModel directly to any LLM client to receive validated, type-safe Python objects.

  • PydanticResponseParser: Auto-generates structural instructions and validates output.
  • JSONResponseParser: Robust regex-based dictionary extraction.
  • XMLResponseParser: Flat dictionary extraction from XML tags.

Fixes & Improvements

  • LLM Clients: Updated default Gemini model to gemini-2.5-lite.
  • API Key Detection: Fixed a race condition where environment variables were not picked up correctly at import time.
  • MRO Fixes: Resolved method resolution order conflicts in the parser inheritance hierarchy.

[v0.1.5] - 2026-04-12

Major Additions

OpenAI Integration

We've added a robust OpenAIClient that leverages the modern Responses API. This ensures compatibility with the latest reasoning models (o-series) and provides detailed metadata like cached tokens.

Automated Documentation

The library now features a professional documentation site built with MkDocs. This includes an automated API reference that stays in sync with the source code.

Comprehensive Examples

A new examples/ directory has been populated with scripts covering everything from basic chat to complex parallel execution graphs.

Fixes & Improvements

  • MaticGraph: Fixed a critical constructor assignment bug and resolved Unicode encoding issues on Windows.
  • LLM Clients: Better API key validation and improved fallback logic for Gemini.
  • Aesthetics: Replaced library-wide emojis with clean, professional Material Design icons.

[v0.1.4] - 2025-10-24

Added

  • Parallel execution support in MaticGraph: Explicit fan-out/fan-in patterns using the parallel_group() method.
  • System instruction support: Properly formatted system instructions for Google Gemini models.

Changed

  • Async Client Management: Better resource cleanup and timeout handling in asynchronous requests.

For full historical details, please consult the CHANGELOG.md in the repository.