Why should you care?

Oversized chunks retrieve irrelevant text; undersized ones lose context — chunk strategy drives RAG accuracy.

See it live — copy this example

Paste into an ASP.NET Core 8+ / AgentNest project, then run with dotnet run (set your API keys in user-secrets).

// AgentNest.Rag/Chunking/MarkdownChunker.cs
public static class MarkdownChunker
{
    public static IEnumerable<KnowledgeChunk> Split(string docId, string markdown, int maxChars = 1200)
    {
        var sections = markdown.Split("\n## ", StringSplitOptions.RemoveEmptyEntries);
        foreach (var (section, i) in sections.Select((s, i) => (s, i)))
        {
            for (var start = 0; start < section.Length; start += maxChars)
            {
                var slice = section.Substring(start, Math.Min(maxChars, section.Length - start));
                yield return new KnowledgeChunk($"{docId}:{i}:{start}", slice.Trim(), docId);
            }
        }
    }
}

What happened?

• MarkdownChunker splits on H2 headings then caps slice length.
• Each KnowledgeChunk gets a stable Id for vector upsert.

Practice next

1. Add MarkdownChunker under AgentNest.Rag/Chunking.
2. Tune maxChars per content type — clinical vs CRM.
3. Preserve metadata: source file, page, section title.
4. Add 100-character overlap between consecutive slices.
5. Use Semantic Kernel TextChunker for PDF pipelines.

Remember

Chunk documents before embedding in AgentNest RAG. Use structure-aware splits for manuals and policies. Attach stable IDs and source metadata to every chunk.

ERP manual ingestion

Finance uploads 800-page GL policy PDF.

Outcome: MarkdownChunker produces retrievable sections cited in variance explanations.