v0.3.2 - CancellationToken Support

What's New

CancellationToken Support

Added cooperative cancellation throughout the query pipeline, making it easy to cancel in-flight LLM requests - useful for GUI applications where users may want to stop a query mid-flight.

Changes:

• RagEngine.QueryAsync now accepts an optional CancellationToken
• RetrievalService.QueryAsync propagates the token through the pipeline
• LlmGenerator.GenerateAnswerAsync and the internal CallLlmAsync pass the token to all HttpClient calls
• OperationCanceledException is re-thrown (not swallowed by the error handler), so callers receive clean
  cancellation semantics

Usage:

using var cts = new CancellationTokenSource();

// Cancel after 30 seconds, or call cts.Cancel() from a button click, etc.
cts.CancelAfter(TimeSpan.FromSeconds(30));

var answer = await ragEngine.QueryAsync("What is RAG?", cancellationToken: cts.Token);

This is a non-breaking change — the parameter is optional with a default of CancellationToken.default.

v0.3.1 - Thinking Model Compatibility & Documentation

Bug Fix

Wildcard query crash with reasoning/thinking models (QueryExpander, HybridRetriever)

Models that emit chain-of-thought output (DeepSeek-R1, Qwen3, etc.) sometimes prefix their expanded query variants with markdown list markers (* , - , 1. ). When those markers survived into the Lucene BM25 query parser, it threw a ParseException because * and ? are illegal as the first character in a wildcard query.

Two-layer fix:

• QueryExpander now strips leading bullet and numbered-list markers from each expanded variant before they enter the pipeline.
• HybridRetriever catches any remaining ParseException and retries with the query text fully escaped (retrieval degrades gracefully rather than crashing).

Improvements

XML documentation on all public members

All publicly visible types and members now carry <summary> XML doc comments, eliminating the CS1591 warning flood when building with <GenerateDocumentationFile>true</GenerateDocumentationFile>. This also means IntelliSense tooltips are populated when consuming PyRagix.Net as a library.

v0.3.0 - OpenAI-compatible LLM endpoint

What's New

OpenAI-compatible LLM endpoint

PyRagix.Net now works with any local inference server that implements the OpenAI /v1/chat/completions API - KoboldCpp, llamacpp, LM Studio, vLLM, LocalAI, and Ollama (via its /v1 endpoint).

The Ollama-native /api/generate wire format has been replaced with the OpenAI chat completions format. Health checks now use /v1/models.

Breaking: config key rename

Update your settings.toml:

Old key	New key
OllamaEndpoint	LlmEndpoint
OllamaModel	LlmModel
OllamaTimeout	LlmTimeout

Default port changed from 11434 to 8080 (llamacpp default).
Set LlmEndpoint to your server's base URL - do not include /v1/ in the value.

LlmEndpoint = "http://localhost:5001"   # KoboldCpp
LlmEndpoint = "http://localhost:8080"   # llamacpp
LlmEndpoint = "http://localhost:11434"  # Ollama
LlmModel    = "your-model-name"

Other fixes

• Console app now loads settings.toml from the working directory (run from pyragix-net-console/)
• Errors when the LLM server is unreachable now exit cleanly without a stack trace
• Removed unused Microsoft.SemanticKernel dependency (carried a CVE)
• Updated ONNX Runtime, EF Core, AngleSharp, and test packages to latest

Upgrading from v0.2.0

1. Rename the three config keys above in your settings.toml
2. Verify LlmEndpoint does not include a /v1/ suffix

v0.2.0

What's New

--fresh flag

Ingest command now accepts --fresh to wipe existing indexes and database before rebuilding from scratch. Previously required manual deletion of artifacts.

dotnet run -- ingest ./docs --fresh

ONNX Execution Provider

Replaced the GpuEnabled bool with ExecutionProviderPreference, a three-way enum controlling how ONNX Runtime picks a compute device for embedding and reranking inference:

Value	Behaviour
Cpu	CPU only. Default.
Auto	Try CUDA, fall back to CPU silently.
Cuda	Require CUDA — fails at startup if unavailable.

ExecutionProviderPreference = "Auto"
GpuDeviceId = 0

Settings format fixed

settings.toml / settings.example.toml were copies of the Python project's sectioned format and were silently ignored by the C# config binder. Converted to flat PascalCase format matching PyRagixConfig property names.

Bug Fixes

• Fixed CI workflow failing on dotnet tool install due to env: PATH: not expanding shell variables in GitHub Actions YAML.
• Fixed reportgenerator step treating ; in -reporttypes: as a shell command separator.

Initial Release

PyRagix.Net v0.1.0 Release Notes

Highlights

• Local-first Retrieval-Augmented Generation engine targeting .NET 9.0 with a shared codebase for both library (pyragix-net/) and demo console (pyragix-net-console/).
• End-to-end RAG pipeline: multi-format ingestion (PDF, HTML, image OCR), semantic chunking, ONNX-based embeddings, FAISS/Lucene hybrid retrieval, cross-encoder reranking, and Ollama-driven answer generation.
• Configurable via TOML (pyragix-net/settings.toml) or programmatic PyRagixConfig, including knobs for query expansion, hybrid weighting, reranking depth, GPU enablement, and batching.
• Works fully offline once ONNX models and Ollama weights are exported locally—no remote APIs or cloud services required.

What's in v0.1.0

• RagEngine public API encapsulating ingestion and retrieval orchestration with EF Core metadata persistence.
• Document pipeline featuring PdfPig, AngleSharp, and Tesseract OCR, plus sentence-aware SemanticChunker with overlap controls for higher recall.
• Hybrid retrieval stack combining FAISS (or managed inner-product fallback) with Lucene BM25 and SQLite metadata for Reciprocal Rank Fusion.
• Retrieval refinements including multi-query expansion via Ollama, ONNX Runtime cross-encoder reranking, and streaming-friendly answer generation.
• Console CLI (dotnet run -- ingest … / -- query …) that exercises the full pipeline using the shared configuration file.
• Docs & tooling covering ONNX export (docs/ONNX_SETUP.md), technology explainer, OS-specific behavior, and test harness instructions.

Recent Improvements

• Reranker honors caller limits – Reranker now respects the requested topK, enabling tighter answer contexts and lower latency (4f5638e).
• End-to-end resiliency – Polly-based retries guard ingestion, retrieval, and Ollama calls against transient IO or HTTP faults (6c54679).
• Startup validation – Pipeline bootstrapping checks for required models, tessdata, and database paths before ingesting or querying (617f623).
• Cross-platform vector index – Added managed inner-product index fallback plus docs on when to regenerate FAISS files when switching OSes (1a52419, f26ddd9).
• Operator ergonomics – Structured ingestion progress output and clarified README/asset guidance; build scripts now quote arguments and ensure bundled dotnet tools resolve (71e8658, 721658a, fef4d22, a6a9ef8).

Setup Essentials

1. Models – Export embedding (sentence-transformers/all-MiniLM-L6-v2) and reranker (cross-encoder/ms-marco-MiniLM-L-6-v2) ONNX bundles into pyragix-net-console/Models/... (see docs/ONNX_SETUP.md).
2. Config – Copy pyragix-net/settings.example.toml → settings.toml, point to ONNX paths, database file, and Ollama endpoint/model.
3. Artifacts – Keep pyragix.db, faiss_index.bin, and lucene_index/ alongside the process working directory; delete them when switching operating systems before re-ingesting.
4. Runtime deps – Install Tesseract eng.traineddata under ./tessdata/ for OCR, ensure Ollama is running locally, and restore/build with dotnet restore && dotnet build.

Happy querying!