Cisharpai.Rag 0.4.0

Cisharpai.Rag

Provider-independent RAG ingestion foundations for .NET 8 and .NET 10: fixed-size text chunking, bulk float embeddings with token-aware batching, bounded concurrency and retry, and a composable document pipeline using IEmbeddingClient.

using Cisharpai.Rag;
using Cisharpai.Rag.Chunking;
using Cisharpai.Rag.Embeddings;
using Cisharpai.Rag.Models;

// embeddingClient is an existing configured Cisharpai IEmbeddingClient.
var pipeline = new RagIngestionPipeline(
    new FixedSizeChunker(new FixedSizeChunkerOptions
    {
        ChunkSize = 1024,
        Overlap = 128
    }),
    new BulkEmbeddingProcessor(embeddingClient, new BulkEmbeddingOptions
    {
        // Provider presets set MaxBatchItems/MaxBatchTokens; apply before your own overrides.
        Model = "text-embedding-3-small",
        MaxConcurrency = 3
    }.ApplyProfile(EmbeddingProviderProfile.OpenAi)));

await foreach (var batch in pipeline.IngestAsync(new[]
{
    new RagDocument("guide", "Your document text")
}))
{
    if (!batch.IsSuccess)
    {
        Console.Error.WriteLine($"Batch {batch.BatchIndex} failed: {batch.ErrorMessage}");
        continue; // a failed batch does not stop the run
    }
    foreach (var item in batch.Items)
        Console.WriteLine($"{item.Chunk.DocumentId}:{item.Chunk.Index}: {item.Vector.Length}");
}

Register a provider separately, then use services.AddCisharpaiRag(options => ...) for DI. A factory overload selects keyed embedding clients. The same pipeline accepts IAsyncEnumerable<RagDocument> and cancellation; BulkEmbeddingProcessor also accepts existing chunk streams.

Defaults: chunk size 1024 Unicode scalar values, overlap 128, 32 items per batch with no token budget, sequential requests (MaxConcurrency = 1), 3 retries on transient failures, document input type, float vectors. BulkEmbeddingOptions.ForProvider/ApplyProfile set per-provider batch ceilings (OpenAi, AzureOpenAi, AzureAiInference, Cohere, Conservative). Chunk positions use UTF-16 offsets. Options are validated and snapshotted at construction/resolution. A failed batch is reported through EmbeddingBatchResult and the run continues; IProgress<BulkEmbeddingProgress> reports how far along ingestion is. Read-ahead is bounded by MaxPendingBatches, so memory does not grow with corpus size. Parsing, vector storage, retrieval, and generation are outside this package. For real token counting (replacing the default s.Length / 4 heuristic), install Cisharpai.Rag.Tokenizers and wire in TiktokenCounter.ToTokenEstimator().

See the RAG usage guide for host configuration, keyed providers, option tables, streaming, and failure handling.