RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM

RAGU, developed by researchers at ITMO University and partner institutions, is an open-source GraphRAG engine designed to reduce the noise and brittleness of single-pass knowledge-graph construction used in systems such as Microsoft GraphRAG and LightRAG. Its six-stage pipeline separates typed entity extraction from relation extraction under the NEREL schema, then applies DBSCAN-based deduplication, LLM summarization, Leiden community detection, and configurable local, global, hybrid, or query-plan search. The companion Meno-Lite-0.1 is a compact 7B model fine-tuned for contextual comprehension and extraction rather than broad factual recall; it surpasses Qwen2.5-32B by 12.5% relative harmonic mean on knowledge-graph construction while matching larger models on GraphRAG tasks. On GraphRAG-Bench Medical, RAGU achieved evidence recall up to 0.84 versus ≤0.76 for competitors, and on Creative Generation it reached 59.0 answer correctness and 57.4 coverage, compared with HippoRAG 2’s 56.9 and 34.7. HippoRAG 2 remained stronger for precise single-fact retrieval and the hardest MuSiQue multi-hop benchmark, while RAGU’s apparent deficits on other factoid tests narrowed substantially under terse answer formatting. Unlike pipelines requiring large hosted models such as OpenAI’s GPT-4-class APIs, RAGU runs on a single consumer GPU, processes roughly 8K tokens per document at about 2K tokens per second, and costs approximately $0.001 per document versus $0.10 for commercial indexing. The implementation emphasizes production reliability through swappable graph, key-value, and vector stores, Pydantic-validated outputs, asynchronous execution, incremental updates, deterministic identifiers, and approximately 374 automated tests.