Skip to content

Changelog

All notable changes to this project will be documented in this file. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

...

[0.0.27] - 2025-06-17

Added

  • Update batch_retrieve for RAGSystems to use batch_retrieve of Knowledge Stores if implemented (#441)
  • Implement batch_retrieve for Qdrant sync/async knowledge stores (#439)
  • Add batch_retrieve to KnowledgeStore classes that raise NotImplementedError by default (#436)
  • Add batch methods for RAGSystem (#270)
  • AsyncQdrantKnowledgeStore (#430)

[0.0.26] - 2025-06-10

Changed

  • Batch execution of cosine sim for InMemoryKnowledgeStore (#366)

[0.0.25] - 2025-06-09

Changed

  • pre-commit hook for mypy with pass_filenames = False to resolve error with .pyi and .py files in same dir (#419)

Added

  • Lazy loading of heavy classes and modules to improve import time of top-level module (#419)
  • Fix monitor_live method (#413)
  • Add ProcessMonitor for notebook utils (#412)
  • Add docs for the LangChain Bridge

[0.0.24] - 2025-06-02

Added

  • [Feature] Add to_sync() methods for AsyncKnowledgeStore and AsyncRAGSystem and their NoEncode variants. (#404)
  • Add rerank_callback to MCPKnowledgeStore (#402)
  • Add BaseMCPKnowledgeStore (#400)
  • Add MCPKnowledgeStore and MCPKnowledgeSource (#393)
  • Add LangChain bridge (#291)
  • Add AsyncRAGSystem and AsyncNoEncodeRAGSystem (#392)
  • Feature/364 onboard more benchmarks (#391)
  • Validate installed version of bridge framework to confirm compatibility based on BridgeMetadata (#323)
  • Add NoEncodeRAGSystem (#386)
  • Add NoEncodeBaseKnowledgeStore and NoEncodeAsyncBaseKnowledgeStore (#385)

Changed

  • Change RALT and DataCollatorForRALT to allow for union RAGSystem | NoEncodeRAGSystem (#404)
  • Change return type of converter functions to list[KnowledgeNode] and source.retriever() return ~mcp.CallToolResult (#402)
  • Refactor MCPKnowledgeStore and MCPKnowledgeSourceto use new base and have retrieve method (#400)
  • Changed load, count, and persist methods in AsyncNoEncodeRAGSystem to sync (#393)
  • Modified private API for _RAGSystem, _AsyncRAGSystem and no-encode versions (#392)

[0.0.23] - 2025-05-27

Added

  • Add BaseAsyncKnowledgeStore (#262)
  • Add ability to save benchmark evaluations to a JSONL file of BenchmarkEvaluationExamples (#377)

[0.0.22] - 2025-05-26

Added

  • Feature/348 onboard pubmedqa (#363)

Changed

  • Fix bug in HuggingFaceGeneratorMixin.generate() to return newly generated tokens only (#345)
  • Passing kwargs for trainer (#365)

[0.0.21] - 2025-05-23

Changed

  • Fix bug in UnslothFastModelGenerator when calling to_peft and LoRA adapters mismatch dtype with base (#360)
  • Fix bug in HFPreTrainedTokenizer.encode() where shape of raw encoder result tokenizer() is of dimension 2 i.e., batched (#360)
  • Improved DataCollatorForRALT final dtypes after padding is applied (#360)

[0.0.20] - 2025-05-23

Added

  • Support get_peft_model in UnslothFastModelGenerator (#356)
  • New integration: Unsloth 🦥 UnslothFastModelGenerator (#356)

Changed

  • Make knowledge node embedding optional and remove redundant embedding in Qdrant upserts (#353)

[0.0.19] - 2025-05-20

Added

  • Add support for in-memory Qdrant instances (#350)
  • Add EvalError (#339)
  • Add streaming support for HF benchmarks (#337)
  • [Feature] Add HuggingFaceBenchmarkMixin and first HuggingFaceMMLU benchmark (#334)
  • [Feature] Add ExactMatchEvaluationMetric (#333)
  • [Feature] Add base.evals module and BaseBenchmark BaseBenchmarker classes (#326)

Changed

  • Add num_examples to BaseBenchmark (#344)

[0.0.18] - 2025-05-16

Added

  • Trainer managers to root import (#320)

Changed

  • Don't raise MissingExtraError for HF generators at import time (#320)
  • Add public api for fed_rag.trainers (#320)
  • Refactor types to data_structures (#319)

[0.0.17] - 2025-05-16

Changed

  • Rip out public api, as its not standard (#317)

[0.0.16] - 2025-05-16

public API [0.0.2]

  • added retrievers, generators, knowledge stores, and rest of types (#310)

[0.0.15] - 2025-05-16

Added

  • .api public module (#307)
  • New core module that houses RAGSystem (#306)

Changed

  • Refactor: move aux RAG system types to fed_rag.types.rag (#306)

[0.0.14] - 2025-05-14

Added

  • LlamaIndexBridge (#285)
  • Add base bridge mixin (#284)

[0.0.13] - 2025-05-10

Added

  • [Feature] Add Qdrant knowledge store (sync) (#259)

Changed

  • QdrantKnowledgeStore align with qdrant sdk (#273)
  • [Fix] Set timeout param to ~qdrant_client.QdrantClient and use contextmanager for client creation and teardown (#272)
  • add load_nodes_kwargs and unit test QdrantKnowledgeStore (#266)
  • fix count QdrantKnowledgeStore (#265)
  • update federate fine-tune (#256)
  • Move validation of the presence of trainers to BaseTrainerManager (#257)

[0.0.12] - 2025-05-03

Changed

  • Manually handle the padding versus delegating to DataCollatorForCausalLM (#248)
  • Ensure that TrainingArguments.remove_unused_columns is set to False (#248)

[0.0.11] - 2025-05-03

Added

  • HuggingFaceTrainerManager missing implementations for preparing retriever/generator models for training (#245)

Changed

  • [Fix] LSRLoss should have first input in KL div be in log space (#245)
  • [Fix] DataCollatorForLSR should subclass SentenceTransformerDataCollator (#245)
  • [Fix] DataCollatorForLSR should require grads for retriever's for retriever scores, but not for lm scores (#245)

[0.0.10] - 2025-05-02

Added

  • [Feature] Add HuggingFaceTrainerForRALT and associated DataCollatorForRALT (#241)
  • [Feature] Add PyTorchTrainerMixin (#239)
  • add trainers and data collators (#232)
  • [Feature] Add fed_rag.base.data_collators and BaseDataCollator (#231)
  • [Feature] Add target template to DataCollatorForLSR (#230)
  • [Feature] Implement compute_loss for LSRSentenceTransformerTrainer (#229)
  • [Feature] Add HuggingFaceLSRTrainer (#227)
  • [Feature] Adds HuggingFaceTrainerMixin (#226)
  • Add BaseTrainer (#225)
  • [Feature] Add HFRAGTrainer (#220)
  • [Feature] Add PyTorchRAGTrainer (#219)
  • [Feature] Data collators for LSR (both huggingface and torch) (#187)
  • Add exception for missing extra error (#208)

Changed

  • [refactor] Improvements to TrainerManagers and Trainers classes and adds BaseRetrieverTrainer & BaseGeneratorTrainer (#238)
  • [Refactor] Move DataCollatorForLSR to new fed_rag.data_collators.huggingface module (#233)
  • [chore] rename to trainer manager (#223)
  • [chore] Move HFSentenceTransformerRetriever to huggingface module for consistency (#207)

[0.0.9] - 2025-04-25

Added

  • [Feature] Add abstract method BaseGenerator.compute_target_sequence_proba (cherry picked) (#201)
  • [Feature] Add LM Supervised Retriever Loss LSRLoss (#182)

Changed

  • Make huggingface generators more dry (#202)
  • Improve/validate LSRLoss forward interface (#185)

[0.0.8] - 2025-04-09

Added

  • BaseTokenizer.encode now retursn new type EncoderResult (#150)

[0.0.7] - 2025-04-06

Added

  • ManagedInMemoryKnowledgeStore and ManagedMixin for KnowledgeStore [#135]
  • Added name attribute to BaseKnowledgeStore [#135]
  • Knowledge store exceptions [#135]

Changed

  • Persist and load uses name (ks_id only exists for managed version) [#135]
  • Exception modules have their own respective base exception [#135]

[0.0.6] - 2025-03-26

Post

  • Test Zenodo

Added

  • [Feature] Add utils.data module for building RAG Fine-tuning datasets given a RAGSystem and Sequence of examples (#98)
  • [Feature] Adds BaseTokenizer and HFPretrainedTokenizer subclass (#99)
  • [Feature] Enable BaseGenerator and BaseRetriever used in RAGSystem to access the rag system (#95)

Changed

  • Small fixes to problems found while running the quick start example (#93)

Other

  • Update LICENSE (#92)

[0.0.5] - 2025-03-17

Added

  • [Feature] Added HFPeftModelGenerator for efficient fine-tuning (#84)
  • [Feature] Added HuggingFace's peft.PeftModel as an acceptable model type for decoration (#76)
  • Added vector compute submodule (#70)
  • [Example] Added RA-DIT implementation using HFPeftModelGenerator for llama-2-7b (#86)
  • [Example] Added mock generator training loop; federated for a 350M model (#67)

Changed

  • [Example] RA-DIT default generator model to load with device map auto (#88)
  • [Example] RA-DIT re-factor example for better organization (#87)
  • [Feature] Updated set_weights and get_weights for HuggingFaceFlowerClient to include logic for PeftModel (#82)
  • Re-organized RA-DIT example app (#69)

[0.0.4] - 2025-03-14

Added

  • [EXAMPLE] Adds RA-DIT retriever train loop using HuggingFace integrations (#64)
  • Added HuggingFaceFLTask for federated learning with HuggingFace models (#61)
  • Added federate.tester.huggingface and associated inspectors (#59)
  • Added federate.trainer.huggingface decorators (#56)
  • Added ability to specify device on load for HF Models (#55)

[0.0.3] - 2025-03-12

Added

  • Build RAG system implementation (#50)
  • Added RAGSystem class (#45)
  • Implemented InMemoryKnowledgeStore (#43)
  • Added BaseKnowledgeStore and KnowledgeNode models (#41)
  • Added HFSentenceTransformerRetriever (#38)
  • Added BaseRetriever class (#35)
  • Added HFPreTrainedGenerator class for generative models (#34)
  • Implemented BaseGenerator and HFPretrainedModelGenerator classes (#33)
  • Added support for llama3 models (#30)
  • Added example Retrieval-Augmented Generation with LLM integration (#28)
  • Implemented DragonRetriever for example RAG system (#26)

Changed

  • Updated to use HFSentenceTransformerRetriever in examples (#48)

[0.0.2] - 2025-03-01

Changed

  • Quickstart as a workspace member, smaller package builds (#24)

[0.0.1] - 2025-03-01

Added

  • Working QuickStart! (#20)
  • Implementation PyTorchFLTask.server() (#17)
  • PyTorchFLTask and PyTorchFlowerClient (#16)
  • Inspection of tester callable for pytorch.tester decorator (#14)
  • federate.trainer.pytorch decorate inspects trainer loop to extract spec (#12)
  • BaseTaskModel and PyTorchTaskModel (#5)