Is Jev required for answers, or optional?

No. Both Jev reranking and Jev grounding are optional modes alongside heuristic and LLM paths, configured through the grounding provider setting. The core hybrid retrieval and generation run without them.

What happens if the Jev validator goes down?

It fails open: a validator outage never blocks the user and reranking falls back to heuristic order. The digest presents this as deliberate so retrieval stays available.

How does hybrid retrieval handle exact terms?

Vector search finds passages by meaning while a Postgres lexical path catches exact codes, amounts, and names, fused with reciprocal rank fusion plus dedup and coverage repair.

Do prompt or provider changes need a redeploy?

No redeploy. All prompts live in the database and are editable from the admin panel, and the LLM provider switches between configured backends from the same panel.

What do I need to stand it up from source?

JDK 17 to build even though the bytecode target is Java 11, which means the server itself runs on JDK 11 and up, Maven 3.9 or the bundled mvnw, and PostgreSQL 12 or newer with pgvector. Creating the vector extension needs superuser rights. It packages as a WAR, either self-contained or deployed to an external Tomcat 8.5.

Do document embeddings cost me API calls?

No. Embeddings come from BAAI BGE Small EN v1.5 running locally through ONNX at 384 dimensions, so there are no embedding API calls. Ingestion covers Markdown, Word, PDF, PowerPoint, Excel and live web pages, with table-aware chunking for Markdown and Word and deterministic sheet-to-markdown conversion for Excel that spends no LLM quota.

Why can a longer answer be rejected as ungrounded?

The README calls the JEV grounding criteria deliberate complements: a universal quantifier on the true side, every claim supported, paired with an existential one on the false side makes the verdict a function of answer length, so long answers get rejected however well grounded. A test enforces it, and since the judge grades the whole answer, MAX_ANSWER_CHARS must comfortably exceed a typical answer.

CommunityJev Tools & CLIs

Aetheria

Java RAG assistant over org documents with hybrid retrieval and citations, using Jev optionally for chunk reranking and answer-grounding checks.

See highlights

Open github.com

Stars: 1
Language: Java
Published by: Community
Author: zakariaimtiaz
Last verified: Oct 2026

zakariaimtiaz/Aetheria

User repository on GitHub

View on GitHub

Aetheria is an intelligent RAG bot transforming unstructured enterprise data into instant action. Powered by high-performance vector retrieval, LLM synthesis, and **JEV by Typesafe AI** for maximum precision, Aetheria searches technical specs and corporate records to deliver accurate, citation-backed answers in seconds.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A retrieval-augmented knowledge assistant that turns an organisation's documents into cited, verifiable answers.It is an intelligent RAG bot transforming unstructured enterprise data into instant action. Powered by high-performance vector retrieval, LLM synthesis, and JEV by Typesafe AI for maximum precision, Aetheria searches technical specs and corporate records to deliver accurate, citation-backed answers in seconds.

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Evaluation

Used inJev for Software & DevTools

Languages: Java

Highlights

Ingests Markdown, Word, PDF, spreadsheets, and web pages with table-aware chunking and citations.
Fuses vector and lexical retrieval with dedup, diversity caps, and sub-question coverage repair.
Applies Jev Noul judgments per chunk for reranking and for grounding validation when enabled.
Keeps prompts database-driven and providers swappable from an admin panel without restarts.

Watch out

Jev features are optional and fail open by design; confirm grounding behavior on your own documents. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

paper-radar-jev

Python CLI that pulls recent arXiv papers, scores relevance with TypeSafe and writes daily Markdown and JSON rankings with resume-safe storage.