laya-jev-GraphRAG
Graph-database-agnostic GraphRAG framework where swappable Laya or Jev decisions score edges, route intents and gate hallucinations.
Overview
laya-jev-GraphRAG is an agentic GraphRAG framework that sits on top of an existing graph database and replaces generative LLM calls in the pipeline with typed decisions from a swappable System One model: local Laya or cloud Jev. The same four-phase pipeline (ingestion, pre-retrieval, traversal and post-retrieval) runs identically on Neo4j, Memgraph, Apache AGE and Kuzu, with both the database and the decision model selected by one environment variable each.
Three primitives do the work. Score evaluates edges and relationships from 0 to 1 for edge verification at ingestion, as the heuristic in a custom A* traversal, and for context reranking. Noul handles semantic chunking, entity disambiguation, early termination, a hallucination gate and citation verification. Choice handles intent routing, ontology alignment and conflict resolution. The README argues traditional GraphRAG suffers from an LLM call at every hop, unverified hallucinated edges and database lock-in, and that this design addresses all three. Evaluation runs across all four databases.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Score | How valid or relevant is this edge or relationship? | — |
| Yes / no | Is this answer grounded / is this citation supported? (hallucination gate) | — |
| Choice | What is the intent of this query? / Which ontology class fits? | — |
How it works
- Ingest documents with Noul chunking and Score edge verification
- Route the query's intent with Choice
- Traverse the graph with A* using Score as heuristic
- Rerank context and gate hallucinations with Noul
- Return a verified answer with citations
Strengths
- Database-agnostic across four graph stores
- Swappable local or cloud decision model
- Edges verified at ingestion
- No LLM call per hop
Caveats
- 'Production-ready' claim is the author's
- Answer synthesis still needs a generator presumably
- Complex system to deploy
At a glance
- Type
- Service
- Status
- beta
- Language
- Python
- Runtime
- service over Neo4j, Memgraph, Apache AGE or Kuzu
- Has code
- yes
- License
- Apache-2.0
- Stars
- 51
Use cases and integrations
How curated lists describe it
Agentic GraphRAG engine using swappable System One decision models (local Laya / cloud Jev). Features a complete 4-phase pipeline (Ingestion, Pre-Retrieval, Traversal, Post-Retrieval) and evaluation across Neo4j, Memgraph, Apache AGE, and Kùzu driven by a custom A\* traversal algorithm.
logicrw/awesome-jev-projectsA database-agnostic Agentic GraphRAG framework using swappable System One models (local Laya / cloud Jev). A plug-and-play intelligence layer featuring a complete 4-phase pipeline, continuous evaluation and custom A* traversal for any graph database.
wh000wh000/awesome-jev-liveAgentic GraphRAG engine using swappable System One decision models (local Laya / cloud Jev). Features a complete 4-phase pipeline (Ingestion, Pre-Retrieval, Traversal, Post-Retrieval) and evaluation across Neo4j, Memgraph, Apache AGE, and Kùzu driven by a custom A* traversal algorithm.
hellogumbo/awesome-jevAgentic GraphRAG over Neo4j with a swappable decision model, either the Jev API or a local Laya checkpoint.
MrJev/awesome-jev
Frequently asked questions
Which graph databases are supported?
Neo4j, Memgraph, Apache AGE and Kuzu.
Do I need Jev?
No, a local Laya checkpoint can be used instead.
Where are decisions used?
Edge scoring, chunking, intent routing, traversal heuristics, reranking and hallucination gating.
Related projects
Terminal tool that finds link paths between Wikipedia articles: Jev ranks each page's outgoing links while Python runs the search.
ProjectbetaPythonPermission-aware document library using Jev or Cloudflare Clef for hierarchical search, passage scoring and automatic filing.
LibrarybetaTypeScriptPython library that replaces hard-coded RAG thresholds with calibrated decisions for stopping, chunking, selection, abstention and caching.
ServicebetaPythonMilvus-based multi-hop graph RAG library where two-stage Jev Noul scoring selects relations and reranks passages.
LibrarybetaPythonCLI and agent skill that shortlist notes in Obsidian or Logseq locally, then have Jev rerank them, returning original passages.
DemobetaJavaScriptMCP server giving coding agents typed judgment tools (verify, screen, rerank, gate, score) backed by Jev, with auto/review/escalate policy.
MCP serverbetaPython