laya-jev-GraphRAG

Per-hop LLM calls add latency and overflow context, unverified edges compound errors, and single-database wiring blocks swaps. Here Score, Noul and Choice models handle the 19 pipeline steps and the LLM runs once at the end.

Yes. All four implement the same BaseGraphClient interface, so changing GRAPH_DB_BACKEND needs no code changes. The same pipeline and evaluation run identically across them.

Start on Jev for zero-shot accuracy, collect side-by-side Laya versus Jev logs in ablation mode to JSONL, train Laya on them, then flip DECISION_MODEL_BACKEND to laya for local runs.

Yes during ingestion: edge validity scoring prunes hallucinations, disambiguation merges duplicates, and ontology alignment snaps relations to schema before traversal and reranking.

Ingestion scores edge validity, disambiguates entities and aligns relationships; pre-retrieval routes intent and validates seed nodes; the A* traversal scores edges and gates early termination; post-retrieval reranks context, resolves conflicts, gates hallucinations and verifies citations. The generative Llama-3.1-8B runs once at the end.

Laya runs locally at about 33 ms per call on an RTX 5060 in FP16 for roughly 1.2 GB of VRAM and no API cost, while Jev is a cloud call at about 50 ms and $0.042 per million tokens. Jev batches questions into one parallel API call; Laya is GPU-batched.

It runs both Laya and Jev on the same decisions and writes the side-by-side outputs to JSONL, which becomes the RLCD synthetic training data used to fine-tune Laya before you switch DECISION_MODEL_BACKEND to laya.