instinct.cpp provides ready to use alternatives to OpenAI Assistant API and built-in utilities for developing AI Agent applications (RAG, Chatbot, Code interpreter) powered by language models. Call it langchain.cpp if you like.
Repo RobinQu/instinct.cpp, read from GitHub 2 h ago. RobinQu has not joined Take My Repo.
Full Changelog: https://github.com/RobinQu/instinct.cpp/commits/v0.1.5
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.#include <instinct/...>.conan build using Ctest.doc-agent:
retriever-version argument in CLI to control how retriever related components are constructed.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
instinct-assistant: file-search support. It takes a summary guided strategy inspired by RAPTOR, and follows practices adopted by OpenAI. See more at #16instinct-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:
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.file-source metadata field to keep record of file identifier.instinct-data:
Aggregate method on data template classGetObjectState method on IObjectStore.instinct-core:
LambdaInputParsr and LambdaOutputParser to simplify chain implementation.ManagedApplicationContext to have more delicate lifecycle managements.mini-assistant:
instinct-core:
std::optional.RecursiveCharacterTextSplitter with the one in langchain.instinct-server:
HttpLibServer.duckdb to 0.10.28bde9376fa21401cba17e3ae7a094f07fa3e11293baf2cd729a37e3e9eec0868 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.
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.instinct-lab to standalone repo. Benchmarks and experimental code will be moved there.mini-assistant API endpoints.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
mini-assistant binary, which provides OpenAI compatible Assistant API services.instinct-core module
ChronoUtils for datetime related functions.RandomUtils for random number generation.SnowflakeIDGenerator for unique ID generation.instinct-data module
instinct-llm module
OpenAIToolAgentExecutor: Agent executor based on OpenAI function calling API.insitinct-server module
HttpLibServer and DefaultErrorController implementation.HttpLibSession struct for route handler.instinct-llm module
ILanguageModel has BindTools function to support function calling API.instinct-core module
cpptrace.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
IAgentExecutor.ParquetFileIngestor for reading Apache Arrow parquet format.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.RecursiveCharacterTextSplitter, resulting better recall in RAG.rag_evaulation.ipynb for Ollama and vLLM servers. Use mixtral, all-minillm as default models.doc_agent implementation.