👁 Image
Mostlylucid.LucidRAG.DocSummarizer 2.7.5

Mostlylucid.DocSummarizer

NuGet: Mostlylucid.LucidRAG.DocSummarizer

Local-first document summarization library using BERT embeddings, RAG retrieval, and optional LLM synthesis.

Features

• Local-first: Runs entirely offline using ONNX models - no API keys required
• Citation grounding: Every claim is traceable to source segments
• Multiple modes: Pure BERT extraction, hybrid BERT+LLM, full RAG pipeline
• Format support: Markdown, PDF, DOCX, HTML, URLs
• Vector storage: In-memory, DuckDB (embedded), or Qdrant (external)
• Multi-framework: .NET 8, .NET 9, and .NET 10 support
• OpenTelemetry: Built-in distributed tracing and metrics
• Resilience: Polly-based retry, circuit breaker, and rate limiting

Installation

dotnet add package Mostlylucid.LucidRAG.DocSummarizer

Quick Start

// Register in DI
builder.Services.AddDocSummarizer();

// Inject and use
public class MyService(IDocumentSummarizer summarizer)
{
 public async Task<string> GetSummaryAsync(string markdown)
 {
 var result = await summarizer.SummarizeMarkdownAsync(markdown);
 return result.ExecutiveSummary;
 }
}

Configuration

Basic Configuration

builder.Services.AddDocSummarizer(options =>
{
 // Use local ONNX embeddings (default, no external services)
 options.EmbeddingBackend = EmbeddingBackend.Onnx;

 // Or use Ollama for embeddings
 options.EmbeddingBackend = EmbeddingBackend.Ollama;
 options.Ollama.BaseUrl = "http://localhost:11434";
 options.Ollama.EmbedModel = "nomic-embed-text";
});

From Configuration File

{
 "DocSummarizer": {
 "EmbeddingBackend": "Onnx",
 "BertRag": {
 "VectorStore": "DuckDB",
 "ReindexOnStartup": false,
 "CollectionName": "my-documents"
 },
 "Onnx": {
 "EmbeddingModel": "AllMiniLmL6V2"
 }
 }
}

builder.Services.AddDocSummarizer(
 builder.Configuration.GetSection("DocSummarizer"));

Embedding Models

Summarization Modes

// Pure BERT - no LLM needed, fastest
var summary = await summarizer.SummarizeMarkdownAsync(
 markdown,
 mode: SummarizationMode.Bert);

// BertRag - full pipeline with LLM synthesis
var summary = await summarizer.SummarizeMarkdownAsync(
 markdown,
 focusQuery: "What are the key architectural decisions?",
 mode: SummarizationMode.BertRag);

Vector Store Backends

// In-memory (no persistence, fastest)
options.BertRag.VectorStore = VectorStoreBackend.InMemory;

// DuckDB (embedded file-based, default)
options.BertRag.VectorStore = VectorStoreBackend.DuckDB;

// Qdrant (external server, best for production)
options.BertRag.VectorStore = VectorStoreBackend.Qdrant;
options.Qdrant.Host = "localhost";
options.Qdrant.Port = 6334;

Output Models

DocumentSummary

record DocumentSummary(
 string ExecutiveSummary, // Main summary text
 List<TopicSummary> TopicSummaries, // Topic-by-topic breakdown
 List<string> OpenQuestions, // Questions that couldn't be answered
 SummarizationTrace Trace, // Processing metadata
 ExtractedEntities? Entities); // Named entities (people, places, etc.)

Query Mode

Ask questions about documents with evidence-grounded answers:

var answer = await summarizer.QueryAsync(
 markdown: documentContent,
 question: "What database technology is recommended?");

Console.WriteLine(answer.Answer);
Console.WriteLine($"Confidence: {answer.Confidence}");

foreach (var evidence in answer.Evidence)
{
 Console.WriteLine($" [{evidence.SegmentId}] {evidence.Text}");
}

Segment Extraction

Extract segments without summarizing - useful for building search indexes:

var extraction = await summarizer.ExtractSegmentsAsync(markdown);

foreach (var segment in extraction.TopBySalience)
{
 Console.WriteLine($"[{segment.Type}] {segment.Text}");
 Console.WriteLine($" Salience: {segment.SalienceScore:F2}");
}

OpenTelemetry Observability

The library includes built-in OpenTelemetry instrumentation for distributed tracing and metrics. All instrumentation follows OpenTelemetry semantic conventions.

Activity Sources (Distributed Tracing)

Each activity includes relevant tags (e.g., url.host, http.response.status_code, error.type) for filtering and analysis.

Metrics

DocumentSummarizerService Metrics

OllamaService Metrics

WebFetcher Metrics

Metric Dimensions (Tags)

Common tags available on metrics:

Example: Wire up in ASP.NET Core

builder.Services.AddOpenTelemetry()
 .WithTracing(tracing => tracing
 .AddSource("Mostlylucid.DocSummarizer")
 .AddSource("Mostlylucid.DocSummarizer.Ollama")
 .AddSource("Mostlylucid.DocSummarizer.WebFetcher")
 .AddOtlpExporter())
 .WithMetrics(metrics => metrics
 .AddMeter("Mostlylucid.DocSummarizer")
 .AddMeter("Mostlylucid.DocSummarizer.Ollama")
 .AddMeter("Mostlylucid.DocSummarizer.WebFetcher")
 .AddOtlpExporter());

Example: Console Application

using var tracerProvider = Sdk.CreateTracerProviderBuilder()
 .AddSource("Mostlylucid.DocSummarizer")
 .AddSource("Mostlylucid.DocSummarizer.Ollama")
 .AddSource("Mostlylucid.DocSummarizer.WebFetcher")
 .AddConsoleExporter()
 .Build();

using var meterProvider = Sdk.CreateMeterProviderBuilder()
 .AddMeter("Mostlylucid.DocSummarizer")
 .AddMeter("Mostlylucid.DocSummarizer.Ollama")
 .AddMeter("Mostlylucid.DocSummarizer.WebFetcher")
 .AddConsoleExporter()
 .Build();

Grafana/Jaeger Integration

Send telemetry to Jaeger or Grafana Tempo:

builder.Services.AddOpenTelemetry()
 .WithTracing(tracing => tracing
 .AddSource("Mostlylucid.DocSummarizer")
 .AddSource("Mostlylucid.DocSummarizer.Ollama")
 .AddSource("Mostlylucid.DocSummarizer.WebFetcher")
 .AddOtlpExporter(o => o.Endpoint = new Uri("http://jaeger:4317")))
 .WithMetrics(metrics => metrics
 .AddMeter("Mostlylucid.DocSummarizer")
 .AddMeter("Mostlylucid.DocSummarizer.Ollama")
 .AddMeter("Mostlylucid.DocSummarizer.WebFetcher")
 .AddPrometheusExporter()); // For Prometheus/Grafana

HTTP Resilience

The WebFetcher service includes Polly-based resilience:

• Circuit Breaker: Opens after 5 failures in 30s, stays open for 60s
• Retry: Exponential backoff with jitter (3 attempts, 0.5s-30s)
• Rate Limiting: Respects HTTP 429 Retry-After headers
• Permanent Failures: 403/404/401 fail immediately with clear messages

Dependencies

• Supported: .NET 8.0+, .NET 9.0+, .NET 10.0+
• Included: ONNX Runtime, Markdig, PdfPig, OpenXml, AngleSharp, Polly, OpenTelemetry.Api
• Optional: Ollama (for LLM synthesis), Docling (for complex PDF conversion)

License

MIT