Full Changelog: https://github.com/RobinQu/instinct.cpp/commits/v0.1.5

• Features
  • instinct-transformer: New bge-m3 embedding model. Generally speaking, bge-reranker and bge-embedding are still in preview as they are not fast enough for production.
  • instinct-llm: New JinaRerankerModel for Reranker model API from Jina.ai.
  • instinct-retrieval: New DuckDBBM25Retriever for BM25 keyword based retriever using DuckDB's built-in function.
• Improvements
  • Move example code to standalone repository: instinct-cpp-examples.
  • Rename for all files for camel-case naming conventions
  • Build system:
    • Fix include paths for internal header files. Now all files are referenced using angle bracket pattern like #include <instinct/...>.
    • Rewrite Cmake install rules.
    • Run unit tests during conan build using Ctest.
  • doc-agent:
    • Use retriever-version argument in CLI to control how retriever related components are constructed.
    • Rewrite lifecycle control using application context
  • instinct-retrieval: Fix RAG evaluation. RAG pipeline with MultiPathRetriever should get score more than 80%.

Full Changelog: https://github.com/RobinQu/instinct.cpp/commits/v0.1.4

• Features
  • instinct-assistant: file-search support. It takes a summary guided strategy inspired by RAPTOR, and follows practices adopted by OpenAI. See more at #16
    . It has some limitations, please refer to #22.
  • new instinct-transfomer module that ships built-in models based on transformer. Currently, only BGE-M3-Reranker is implemented. Most code in first version implementation is refined (copied) from chatllm.cpp for quick start.
  • instinct-retrieval:
    • a new MultiPathRetriever that handle recalls from multiple child retrievers and score results with a reranker model.
    • SimpleRetrieverOperator and DuckDBVectorStoreOpeartor for multi-instance managements.
    • CitationAnnotatingChain and SummaryChain to support citations and summary required in instinct-assistant.
    • Added a mandatory file-source metadata field to keep record of file identifier.
  • instinct-data:
    • Aggregate method on data template class
    • Added GetObjectState method on IObjectStore.
  • instinct-core:
    • Added LambdaInputParsr and LambdaOutputParser to simplify chain implementation.
    • ManagedApplicationContext to have more delicate lifecycle managements.
• Improvements
  • mini-assistant:
    • many bug fixes in related classes.
    • CLI options: we can now assign model provider separately for embedding model and chat model.
  • instinct-core:
    • Make options on chat model and embedding model optional using std::optional.
    • Fix a consistency issue of RecursiveCharacterTextSplitter with the one in langchain.
    • Better control of preloading assets of tokenizers.
  • instinct-server:
    • fix graceful shutdown of HttpLibServer.
  • bump duckdb to 0.10.2

8bde9376fa21401cba17e3ae7a094f07fa3e11293baf2cd729a37e3e9eec0868  doc-agent_linux-amd64.tar.gz
8fad76f8c402240189817b221442b7a596b927653e36ae61d431d1b0cd3776ef  doc-agent_macos-arm64.tar.gz
8042950bc8a28df375e58589423d87c77bae545d4f71d5acf1e224932cfae32f  mini-assistant_linux-amd64.tar.gz
848d1e7a4a2b80c42a7348d37d0337788e67f13d745acbed3652d65b964164c8  mini-assistant_macos-arm64.tar.gz

Full Changelog: https://github.com/RobinQu/instinct.cpp/commits/v0.1.3

Sorry for the delay. Hacking with new agent executor costs me more time than expected. But it's fun and enlightening. As a result, work on rerank models will delay to next release.

• Features
  • New LLMCompilerAgentExecutor as default executor in mini-assistant. Please see this notebook for evaluation results.
  • POST /v1/threads/runs endpoint, which is missed in last release, is added.
  • xn::steps::branch and BranchStepFunction for binary branch composing.
• Improvements
  • Migrate instinct-lab to standalone repo. Benchmarks and experimental code will be moved there.
  • Many fixes for typos in mini-assistant API endpoints.
  • Add more trace logs in debug mode.
  • BaseChatModel and BaseLLM supports runtime configuration with configure(const ModelOverrides&) method.
  • HttpLibServer defaults to listen 0.0.0.0.

SHA256 digest for assests:

e0ebb75722e6eedba6c60b01a1a1938e15bd1ea043949c0e8cc97ee60fedc099  mini-assistant_linux-amd64.tar.gz
114a04e95dee7b51559fe78003c78dfd996ee2d5e035e5881f6fa16f9a6acf25  mini-assistant_macos-arm64.tar.gz
9db7fb3513b0546ce2cfaea8871a9df7437d6cbf8e8ac57edf25b4283b656cca  doc-agent_linux-amd64.tar.gz
90385ddb9f3e646eda3b05c0e9b46af4cbb5b49b598bb391dff322d6172097d1  doc-agent_macos-arm64.tar.gz

Full Changelog: https://github.com/RobinQu/instinct.cpp/commits/v0.1.2

• Features
  • First release of mini-assistant binary, which provides OpenAI compatible Assistant API services.
  • instinct.cpp library:
    • instinct-core module
      • ChronoUtils for datetime related functions.
      • RandomUtils for random number generation.
      • SnowflakeIDGenerator for unique ID generation.
    • instinct-data module
      • Data mapper classes that is built for simple ORM, with protobuf message.
      • Task scheduler abstraction and in-process task queue implementation.
      • Object store abstraction and local filesystem implementation.
    • instinct-llm module
      • OpenAIToolAgentExecutor: Agent executor based on OpenAI function calling API.
    • insitinct-server module
      • Error controller for HttpLibServer and DefaultErrorController implementation.
      • HttpLibSession struct for route handler.
• Improvements
  • instinct-llm module
    • function call protocol is updated to match OpenAI's standard.
    • ILanguageModel has BindTools function to support function calling API.
  • instinct-core module
    • Exceptions are now with stacktrace generated by cpptrace.
    • assertion errors now throw ClientException which is sublcass of InstinctException.

SHA256 for assets:

cff7ddc64c37aa1fbe7a772a99ab6461f1d8cd93f802897d586586cc74b1f45d  mini-assistant_macos-arm64.tar.gz
73d265ea05aa52d2bd1325c21965ad2accb629fbe218a8ba2f240b88f4a19967  mini-assistant_linux-amd64.tar.gz
872157a97a25551ed4d5768f64018aa38f6a7e3b9516286e2592687744e32bf2  doc-agent_linux-amd64.tar.gz
79c4c7b178bafa560a4a8f7721d8f8361f35e624240972d793577687ecb49368  doc-agent_macos-arm64.tar.gz

Full Changelog: https://github.com/RobinQu/instinct.cpp/commits/v0.1.1

• Features
  • Initial agent frameworks. See IAgentExecutor.
  • Function calling: ReACT style tool agent is implemented as intial support for LLM agent. See TestReactAgent.
  • Ingestor:
    • ParquetFileIngestor for reading Apache Arrow parquet format.
  • Utitlies:
    • HashUtils for MD5, SHA1, SHA256 hash functions.
    • IOUtils, SystemUtils are added.
    • ExecuteWithCallback and ExecuteBatch are added to IHttpClient.hpp.
    • IFileVault and FileSystemFileVault for advanced file resource management.
• Improvements
  • bugfixes in RecursiveCharacterTextSplitter, resulting better recall in RAG.
  • Rewrite HTTP server classes with more modular controller design.
  • Code cleanup for DuckDB store classes.
  • Adapt rag_evaulation.ipynb for Ollama and vLLM servers. Use mixtral, all-minillm as default models.
• Known issues
  • CI is not ready and some tests relies on local setup. You may experience problems for building and running on your machines.
  • Large PDF and long text are not tested with current doc_agent implementation.