3GPP-Everything
A retrieval-augmented generation agent over the complete 3GPP Release-18 and Release-19 5G specification set, answering natural-language questions with paragraph-level citations back to the original text and strict grounding.
3GPP-Everything is a production RAG system covering all Release-18 and Release-19 5G technical specifications published by GSMA -- 1,270 documents split into roughly 395,000 passage chunks. It answers natural-language questions about the specifications with inline citations that link back to the exact section, and refuses to answer when no supporting passage is found rather than relying on the model's general knowledge.
Retrieval combines dense Qdrant vector search with BM25 sparse search through reciprocal rank fusion, followed by reranking and a small-to-big expansion that returns the parent section around a matched chunk. Complex, multi-document questions trigger HyDE, multi-query retrieval, and a self-RAG loop that scores its own grounding, coverage, and confidence and retries before answering.
The system is built on LangGraph for orchestration with a PostgreSQL checkpointer, a custom async retrieval layer over Qdrant and BM25 rather than a high-level framework retriever, and a self-hosted LiteLLM proxy so the underlying model can be swapped freely.