Add self-correcting RAG agent notebook with LangGraph

[LuciferDono]

Summary

Adds a new notebook notebooks/llm-agent-langgraph/llm-agent-langgraph.ipynb implementing a self-correcting RAG agent using LangGraph and OpenVINO.

The notebook demonstrates a stateful RAG pipeline that goes beyond linear retrieve-then-generate by adding:

• Document grading — LLM evaluates relevance of each retrieved document
• Query rewriting — if no relevant documents found, the agent reformulates the query and re-retrieves
• Conditional routing via LangGraph state machine with a configurable max-rewrite limit

Key features

• Model selection widget (Phi-3-mini-4k, Llama-3.2-1B, Qwen2.5-1.5B)
• INT4 weight compression via Optimum CLI
• ChromaDB vector store with bge-small-en-v1.5 embeddings
• Interactive Gradio demo showing agent reasoning steps
• Full table of contents, telemetry, and scarf pixel

Files added

• notebooks/llm-agent-langgraph/llm-agent-langgraph.ipynb
• notebooks/llm-agent-langgraph/README.md

Closes #3240

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[aleksandr-mokrov]

@LuciferDono, thanks for the contribution!

I tried running the notebook locally and hit an error during model conversion with optimum-cli. Could you please share the full conversion log from a successful run on your side? Also, which versions of optimum-intel and openvino did you use?

[LuciferDono]

Hey @aleksandr-mokrov, sorry about the wait. I hit the same conversion error on my end and have been debugging it.

Found a few issues in the notebook. Fixing them up and running a clean test now. Will reply here with the full conversion log and the exact openvino/optimum-intel versions once it goes through.

[LuciferDono]

Pushed the fix and tested on Colab (Intel Xeon, Python 3.12). Everything runs through now.

Versions: openvino 2026.1.0, optimum-intel from git HEAD, nncf 3.0.0, transformers 4.53.3, torch 2.8.0+cpu

Main things I changed; switched to cmd_helper.optimum_cli() for the conversion, moved model loading to HuggingFacePipeline.from_model_id with the openvino backend, and pinned torch/torchvision to avoid a version conflict that was breaking the export. Phi-3-mini converts and runs fine, the self-correction loop fires correctly on vague queries.

[LuciferDono]

hey @aleksandr-mokrov, just checking in, did the fix work on your end? let me know if you need anything else from me.

Pushed the fix and tested on Colab (Intel Xeon, Python 3.12). Everything runs through now.

Versions: openvino 2026.1.0, optimum-intel from git HEAD, nncf 3.0.0, transformers 4.53.3, torch 2.8.0+cpu

Main things I changed; switched to cmd_helper.optimum_cli() for the conversion, moved model loading to HuggingFacePipeline.from_model_id with the openvino backend, and pinned torch/torchvision to avoid a version conflict that was breaking the export. Phi-3-mini converts and runs fine, the self-correction loop fires correctly on vague queries.