Awesome AI Agents › Reasoning Frameworks

NirDiamant/Controllable-RAG-Agent

⭐ 1626 Jupyter Notebook repository created 2024-04-03

The Controllable-RAG-Agent project is an advanced Retrieval-Augmented Generation (RAG) solution designed to address complex question answering tasks that go beyond simple semantic similarity retrieval methods. At its core, the project features a sophisticated deterministic graph that functions as the "brain" of a highly controllable autonomous agent. This agent is capable of answering non-trivial questions by leveraging custom datasets, ensuring that responses are grounded solely in the provided data to prevent hallucinations common in AI-generated content. The system operates through a multi-step reasoning process that breaks down complex queries into manageable subtasks. It begins by loading and processing PDF documents, splitting them into chapters, and then cleaning and preprocessing the text for summarization and encoding. Extensive chapter summaries are generated using large language models, and a database of book quotes is created to support specific question answering. The content and summaries are encoded into vector stores for efficient retrieval. When a question is posed, it is anonymized by replacing named entities with variables, allowing the agent to generate a high-level plan to answer the question without bias. This plan is then de-anonymized and broken down into retrievable or answerable tasks. Each task is executed by either retrieving relevant information from the vector stores or generating answers using chain-of-thought reasoning. The agent continuously verifies that generated content is faithful to the original context and adapts its plan based on new information. The final answer is produced by accumulating context and reasoning through the chain-of-thought method. The project also includes performance evaluation using Ragas metrics, which assess answer correctness, faithfulness, relevancy, context recall, and answer similarity. A notable use case demonstrated is the analysis of the first Harry Potter book, showcasing the agent's ability to rely strictly on retrieved information rather than pre-trained knowledge. The project supports integration with various large language models and technologies such as LangChain, FAISS vector stores, and Streamlit for visualization. It offers a comprehensive, controllable, and reliable approach to complex question answering with a focus on minimizing hallucinations and maximizing reasoning transparency.

https://github.com/NirDiamant/Controllable-RAG-Agent

adaptive-planningadvanced-ragagentanonymizationautonomous-agentchain-of-thought-reasoningcomplex-question-answeringcontent-verificationdeterministic-graphfaissgenaihallucination-preventionharry-potter-analysislangchainlanggraphllmllmsmulti-step-reasoningopenaipdf-processingperformance-evaluationpythonragragas-metricsretrieval-augmented-generationstreamlitsummarizationtext-preprocessingvector-store-encoding