Orkeon.Rag.Onnx
1.0.0-rc.4
Prefix Reserved
dotnet add package Orkeon.Rag.Onnx --version 1.0.0-rc.4
NuGet\Install-Package Orkeon.Rag.Onnx -Version 1.0.0-rc.4
<PackageReference Include="Orkeon.Rag.Onnx" Version="1.0.0-rc.4" />
<PackageVersion Include="Orkeon.Rag.Onnx" Version="1.0.0-rc.4" />
<PackageReference Include="Orkeon.Rag.Onnx" />
paket add Orkeon.Rag.Onnx --version 1.0.0-rc.4
#r "nuget: Orkeon.Rag.Onnx, 1.0.0-rc.4"
#:package Orkeon.Rag.Onnx@1.0.0-rc.4
#addin nuget:?package=Orkeon.Rag.Onnx&version=1.0.0-rc.4&prerelease
#tool nuget:?package=Orkeon.Rag.Onnx&version=1.0.0-rc.4&prerelease
Orkeon.Rag.Onnx
License: this package runs cross-encoder/ms-marco-MiniLM-L-6-v2
(Apache-2.0, sentence-transformers / UKP Lab) but ships no weights itself.
ONNX Runtime and Microsoft.ML.Tokenizers are MIT.
Install
dotnet add package Orkeon.Rag.Onnx --prerelease
Published to NuGet.org as part of the consolidated lineup (from
v1.0.0-rc.3; earlier versions reached the GitHub Packages feed only); depends on theOrkeonpackage — see the publication matrix.
Offline story: weights are resolved in this order, lazily at first use —
Orkeon.Rag.Onnx.Modelreferenced → embedded int8 weights (~22 MB), guaranteed offline, zero download. This is the CI path (RAG-04 acceptance criterion 3).OnnxRerankerOptions.ModelPath/VocabPath— VFS virtual paths read throughIFileSystemService. Provision the files withtools/download-reranker-model.sh(SHA-256 verified).- Otherwise → loud, actionable error. The runtime never downloads weights (supply-chain decision: downloading is a host choice, not a library side effect).
What it does
OnnxCrossEncoderReranker : IReranker scores each (query, chunk) pair jointly with the
ms-marco-MiniLM-L-6-v2 cross-encoder — the quality stage of the 50 → 5 cascade
(RerankingOptions.DefaultCandidateK → RerankingOptions.DefaultTopN, guide §7).
Deterministic, CPU, millisecond latency, zero token cost. Scores are sigmoid-mapped
logits in (0, 1) with ScoreOrigin = "cross-encoder".
Tokenization is BERT WordPiece (Microsoft.ML.Tokenizers.BertTokenizer, the model's own
vocab.txt), pairs encoded as [CLS] query [SEP] passage [SEP], batched
(OnnxRerankerOptions.BatchSize, default 16, max sequence 512).
Usage
services.AddOrkeonRag(configuration); // registers the reranker factory (none/noop, llm/listwise)
services.AddOrkeonOnnxReranker(); // contributes onnx/cross-encoder
var reranker = provider.GetRequiredService<RerankerFactory>().Create("onnx");
var top5 = await reranker.RerankAsync(query, candidates, topN: 5, ct);
Factory names: onnx (canonical), cross-encoder (alias).
Native dependency isolation: Microsoft.ML.OnnxRuntime is confined to this opt-in
package, same approach as Orkeon.Tools.Embeddings.Local (plan RAG §7).
| Product | Versions Compatible and additional computed target framework versions. |
|---|---|
| .NET | net10.0 is compatible. net10.0-android was computed. net10.0-browser was computed. net10.0-ios was computed. net10.0-maccatalyst was computed. net10.0-macos was computed. net10.0-tvos was computed. net10.0-windows was computed. |
-
net10.0
- Microsoft.Extensions.Options (>= 10.0.12)
- Microsoft.ML.OnnxRuntime (>= 1.30.0)
- Microsoft.ML.Tokenizers (>= 2.0.0)
- Orkeon (>= 1.0.0-rc.4)
NuGet packages
This package is not used by any NuGet packages.
GitHub repositories
This package is not used by any popular GitHub repositories.
| Version | Downloads | Last Updated |
|---|---|---|
| 1.0.0-rc.4 | 43 | 9/22/2026 |
| 1.0.0-rc.3 | 73 | 9/9/2026 |
Release notes: https://github.com/Orkeon/orkeon/blob/main/CHANGELOG.md