Feature Request : New Notebook for Self-Correcting RAG Agent with LangGraph and OpenVINO

[Ashitpatel001]

The repository's current RAG examples follow a linear execution path. these pipelines work well for straightforward queries, but they are not robust; if the retrieval step retrieves irrelevant documents, the model frequently imagines an answer.

Although loops and self-correction are used in modern "Agentic" workflows to address this, there isn't yet a specific tutorial that shows how to integrate LangGraph with OpenVINO.

Suggested Plan of Action

I suggest creating a new notebook called notebooks/llm-agent-langgraph/llm-agent-langgraph.ipynb.

The goal is to create a Self-Correcting RAG Agent that can grade the relevance of documents it retrieves and, if needed, rewrite its own search queries.

1. State Management: Using LangGraph to maintain the conversation state is one of the key technical concepts.
2. Conditional Logic: Using an OpenVINO-optimized LLM, a "Grader" node is implemented to determine whether the retrieval context is adequate.
3. Cyclic Execution: The agent loops back to rewrite the query and re-retrieve if there is not enough context.

• Orchestration: langgraph, langchain
• Model: Llama-3 or Phi-3 (via optimum-intel / openvino-genai)
• Database: either Chromadb or Faiss

Value to Community

• Demonstrates capability of OpenVINO-optimized models to handle complex, multi-step agentic reasoning (not just single-turn generation).
• Provides a template for developers looking to build "Compound AI Systems" on Intel hardware

I am interested in implementing this notebook. Could you please assign this issue to me? I aim to have a Draft PR ready for review soon.

[Ashitpatel001]

.take

[krishnamurthi-ramesh]

.take