google-deepmind/onetwo v0.4.0

v0.4.0

Repository: google-deepmind/onetwo

Tag: v0.4.0

Published: 2026-04-15T12:44:51Z

Prerelease: no

Release notes:

• Backends
• GoogleGenAIAPI: Set httpx connection limits according to

threadpool_size to allow higher throughput. Handled safety filter blocks more explicitly, raising a specific error. Removed redundant @caching.cache_method decorators from generate_text and chat methods to avoid double caching. Automatically retry on empty responses caused by safety filters. Refactored generate_text and chat to use generate_content as the base implementation. Added support for output_dimensionality and task_type in llm.embed.

• Gemini API: Updated default embedding model name, as

models/embedding-001 is deprecated.

• Core
• Caching: Made SimpleCache.cache_value asynchronous for

non-blocking I/O. Improved printing of cache summaries in colab_utils, including support for TwoLayerCaches and sorting keys. Made cache read resilient to empty file reads caused by race conditions. Added a generic TwoLayerCache implementation.

• Execution & Parallelism: Optimized executing.parallel to reduce

overhead for nested and fixed-size sequences using asyncio.gather directly in a fast path. Allowed changing the default value of max_parallel_executions globally. Propagated context variables in run_method_in_threadpool. Used ContextualExecutor instead of ThreadPoolExecutor to ensure context propagation for debugging tools like Sherlog. Propagated OneTwo tracers across async boundaries in batching.py. Improved exception handling and reporting within the tracing system.

• Agents
• ReAct: Added option retry_on_parsing_error to provide more

detailed error messages on action parsing failures, including the thought and failed action string to help the LLM self-correct.

• Standard library
• Retrieval & QA: Added a comprehensive suite of modules under

third_party/py/onetwo/stdlib/retrieval and third_party/py/onetwo/stdlib/qa to support building advanced Retrieval-Augmented Generation (RAG) and Question Answering systems. Key features include:

• Core Interfaces: Introduced fundamental interfaces for RAG like

Retriever, Index, Searcher, CorpusRewriter, Chunker and DocumentFormatter.

• Data Structures: Added the Document dataclass for representing

content to be indexed and retrieved.

• Indexing: Introduced various index types: in particular

EmbeddingBasedIndex, RewritingIndex.

• CorpusRewriter: CorpusRewriter interface and implementations

to enable flexible document processing pipelines before indexing. Indexes leverage the CorpusRewriter abstraction for more general transformations.

• Chunking: Provided implementations like TextChunker,

NoChunking, and ChunkByMaxTokens for splitting documents.

• Formatter: Provided implementations for formatting documents.
• Searcher: Added BruteForceSearcher for nearest neighbor search

between embeddings.

• Constrained Retrieval: Introduced RetrievalConstraint classes

and the ConstrainedRetriever and ConstrainedSearcher protocols to enable efficient pre-filtering of documents based on metadata fields before vector search.

• Index Serialization: Added EmbeddingBasedIndexState to store

the state of EmbeddingBasedIndex and serializers for persisting and loading index data.

• QA Strategies: Defined interfaces for question answering,

including QAStrategy, ContextualQAStrategy, and RetrievalQAStrategy.

• Colabs
• Introduced a Getting Started tutorial.
• OneTwo tutorial: Demonstrated use of llm.embed() for multimodal

contents.

• Introduced a RAG Tutorial.
• Testing
• Modified LLMForTest tokenizer and added detokenizer for better

testing.

• Evaluation
• Tracing: Improved tracing for metrics in metrics.py and

evaluation.py.

• ot.evaluate: Unified evaluation and agent_evaluation modules

into onetwo.evaluation.evaluation. Made ot.evaluate more generic to support arbitrary example types and strategy signatures. Added dataset_name and dataset_description parameters to evaluate.