Q-RAG-50M-Sovereign
A custom decoder checkpoint for binary relevance classification of a supplied query and passage. The name is a historical repository identifier; it does not identify a complete retrieval service.
Status — October 3, 2026: Qovaryx is on the back burner. Desktop downloads are paused; there is no current public installer, purchase, or trial. These repositories remain as research references. Published weights are separate from desktop product availability.
Intended use
Given a query and a candidate passage in the training-time prompt format, the model predicts a relevance label. It does not create a file index, enumerate the drive, retrieve candidate passages, verify paths, or generate source-grounded answers by itself. Candidate retrieval and output validation are separate components.
What the published reports establish
These are historical measurements, preserved for inspection and not rerun during the October 3 documentation cleanup.
| Report | Scope | Recorded Q-RAG result |
|---|---|---|
| holdout_eval.json | 30 internal query/passage classification rows, exact-match scoring | 30/30 matches. Too small to support a general accuracy claim. |
| benchmark_beir.json | 250 rows from nfcorpus and scifact, including 50 positives | Accuracy 0.896; recorded latency about 168 ms per row in that run. |
| benchmark_vs_embeddings.json | Comparison on the internal holdout | Mixed embedding and reranking baselines, with threshold selection and some failed baselines. |
The BEIR-slice report lists 12 entries including Q-RAG; one baseline failed to load. It is not a complete BEIR evaluation. Several comparison accuracies use accuracy_oracle and a threshold selected on the evaluated slice, so the table is not an independent, uniformly calibrated ranking. In that same run, bge-small-en-v1.5 records accuracy 0.932 at about 38 ms and e5-small-v2 records 0.920 at about 37 ms. The report does not support a blanket claim that Q-RAG beats these models.
The internal holdout and the BEIR slice are different tests. Their results must not be combined into a single general-purpose success rate. The all-correct 30-row fixture does not establish a perfect lower confidence bound on future inputs.
An August 14 artifact-restoration audit verified the identity of the restored published files and treated this checkpoint as a candidate for further evaluation. It did not validate it as a production file retriever. End-to-end retrieval recall, fabricated-path rate, stale-index behavior, and full-system resource use remain unestablished by these reports.
Artifacts and loading
The repository includes pytorch_model.pt, config.json, tokenizer.json, and the evaluation reports linked above. Configuration and tokenizer belong to this checkpoint; the public random-initialized 50M scratch base is a different release.
The historical loader requires a compatible bleeding_edge.model.decoder source package providing FinanceDecoder and DecoderConfig. That source is not included here. A compiled runtime binary does not provide an importable Python package. Inspect the checkpoint format and source dependencies before loading; this is not a standalone Transformers AutoModel checkpoint.
License
Apache-2.0, as recorded in the existing model metadata. This does not establish a blanket license or contamination guarantee for historical training corpora or benchmark datasets.
Project and support
J.E. Herizon LLC, doing business as JE Horizon.