Release v0.19.0

Changed

• Options that were declared but never read now take effect, with their documented defaults. Old → new for a caller
  that sets nothing:
  • Batches run up to 4 items at once (EnableParallelProcessing = true, MaxDegreeOfParallelism = 4, as documented):
    the three evaluators' EvaluateBatchAsync, QAFilterService.FilterAsync/EvaluateBatchAsync,
    ChunkEnrichmentService.EnrichBatchAsync, SummarizeBatchAsync, ExtractKeywordsBatchAsync. They ran one at a time.
    Results keep input order. In a parallel batch the remaining items finish before the first exception is thrown;
    sequential batches still stop at the first failure. QAFilterService asks for three metrics per pair, so it can reach
    12 concurrent model calls. To restore one at a time, set EnableParallelProcessing = false.
  • QA generation prompts carry the difficulty mix (DifficultyDistribution, default about 30 % easy, 50 % medium,
    20 % hard). The line is omitted when all shares are 0.
  • A model intent classification below MinIntentConfidence (default 0.5) is reported as General, keeping the
    measured confidence. Heuristic classifications are never below 0.5.
  • EvaluationOptions.PassThreshold decides MetricResult.IsPassed. The default is now 0.5, which keeps
    IsPassed where it was: it always compared against 0.5, and the declared 0.7 was never read.
• Opt-in options now work:
  • EvaluationOptions.EnableFaithfulness/EnableRelevancy/EnableAnswerability through the new
    QAFilterOptions.Evaluation: a disabled metric is never requested, and its minimum no longer filters the pair.
  • IncludeDetails = false returns scores without Details.
  • ConditionalEnrichmentOptions.DomainGlossary expands terms in the text sent to summarization and keyword
    extraction.
  • QueryPreprocessingOptions.Language asks for keywords and synonyms in that language.
• Breaking — QAPairEvaluation treats a missing score as missing: OverallScore is the mean of the scores
  present (it was null if any was null), and PassesThresholds ignores a null score (it counted as 0). It is false when
  every score is null.
• Breaking — QAFilterService.EvaluateAsync and EvaluateBatchAsync take EvaluationOptions? options before the
  cancellation token. Migration: pass the token by name, or pass null for the options.

Removed

• Breaking — options with nothing behind them: EnrichmentOptions.EnableEntityExtraction and EntityTypes (there
  is no entity extraction step), EvaluationOptions.MaxRetries (nothing in FluxImprover retries; the completion service
  does), QuestionSuggestionOptions.UseConversationHistory and UseDocumentContext (the method you call already
  decides the input). QueryPreprocessingOptions.Language is now string? (default null: no language line) instead of
  "en". Migration: delete the assignments.

Dependencies

• Re-pinned sibling package(s) LMSupply.Generator 0.111.0 -> 0.113.0, LMSupply.Generator.Onnx 0.111.0 -> 0.113.0.
• Re-pinned sibling package(s) LMSupply.Generator 0.113.0 -> 0.115.1, LMSupply.Generator.Onnx 0.113.0 -> 0.115.1.