Benvenuti nella parte 9! Nelle parti precedenti, abbiamo costruito un sistema RAG robusto che elabora i post del blog di markdown e li rende ricercabili attraverso embeddings semantici. Ora è il momento di espandere le nostre capacità con Aggancio - una potente libreria di elaborazione documenti in grado di ingerire PDF, DOCX e altri formati, rendendo il nostro avvocato GPT veramente completo.
NOTA: Questo fa parte dei miei esperimenti con l'AI (elaborazione assistita) + il mio editing. Stessa voce, stesso pragmatismo; solo dita più veloci.
Non che ho bisogno di questo per il blog (è tutto segnato), ma per il completamento ho pensato di mostrare come aggiungere facilmente le capacità di ingestione di documenti alla vostra pipeline RAG. Questo è particolarmente utile se si sta costruendo un sistema che ha bisogno di elaborare documenti legali, contratti, o altri documenti aziendali.
Le pratiche legali moderne si occupano di una varietà di formati di documenti oltre il testo giusto. Gli avvocati devono fare riferimento:
Docling permette al nostro avvocato GPT di elaborare questi formati e renderli ricercabili, creando una base di conoscenze veramente completa che rispecchia come gli studi legali moderni utilizzano AI.
Aggancio è un toolkit di elaborazione di documenti open-source di IBM che:
Docling Serve è il modo più semplice per eseguire Docling come servizio.
# Using the official container image from Quay.io
docker run -p 5001:5001 quay.io/docling-project/docling-serve
# Or with the UI enabled for testing
docker run -p 5001:5001 -e DOCLING_SERVE_ENABLE_UI=1 quay.io/docling-project/docling-serve
Immagini del contenitore disponibili:
| Immagine | Descrizione | Dimensione |
|---|---|---|
quay.io/docling-project/docling-serve |
Immagine di base (Pacchetti PyPI) | ~ 8.7GB (amd64) |
quay.io/docling-project/docling-serve-cpu |
Solo variante CPU | ~4.4GB |
quay.io/docling-project/docling-serve-cu126 |
CUDA 12.6 per GPU | ~10GB |
quay.io/docling-project/docling-serve-cu128 |
CUDA 12.8 per GPU | ~11.4GB |
Endpoint:
http://localhost:5001http://localhost:5001/docshttp://localhost:5001/ui (se abilitato)Per la produzione, aggiungere Docling al tuo esistente docker-compose.yml:
services:
docling:
image: quay.io/docling-project/docling-serve:latest
ports:
- "5001:5001"
environment:
- DOCLING_SERVE_ENABLE_UI=0
- DOCLING_SERVE_MAX_WORKERS=4
volumes:
- docling_cache:/root/.cache
restart: unless-stopped
volumes:
docling_cache:
Prima di integrarsi con C#, testiamo l'API:
# Convert a PDF from URL
curl -X 'POST' \
'http://localhost:5001/v1/convert/source' \
-H 'accept: application/json' \
-H 'Content-Type: application/json' \
-d '{
"sources": [{"kind": "http", "url": "https://arxiv.org/pdf/2501.17887"}],
"options": {
"to_formats": ["md"]
}
}'
# Convert a local file (upload)
curl -X 'POST' \
'http://localhost:5001/v1/convert/file' \
-H 'accept: application/json' \
-F '[email protected]'
Poiché Docling è un servizio Python, ci integreremo tramite HTTP. Costruiamo un robusto client C#.
using System.Text.Json.Serialization;
namespace Mostlylucid.BlogLLM.Core.Models
{
/// <summary>
/// Request to convert documents from URLs or base64 content
/// </summary>
public class DoclingConvertRequest
{
[JsonPropertyName("sources")]
public List<DoclingSource> Sources { get; set; } = new();
[JsonPropertyName("options")]
public DoclingOptions? Options { get; set; }
}
public class DoclingSource
{
[JsonPropertyName("kind")]
public string Kind { get; set; } = "http"; // "http", "base64", "file"
[JsonPropertyName("url")]
public string? Url { get; set; }
[JsonPropertyName("base64")]
public string? Base64Content { get; set; }
[JsonPropertyName("filename")]
public string? Filename { get; set; }
}
public class DoclingOptions
{
[JsonPropertyName("to_formats")]
public List<string> ToFormats { get; set; } = new() { "md" }; // "md", "json", "text"
[JsonPropertyName("ocr")]
public bool Ocr { get; set; } = true;
[JsonPropertyName("table_mode")]
public string TableMode { get; set; } = "accurate"; // "fast", "accurate"
}
/// <summary>
/// Response from Docling conversion
/// </summary>
public class DoclingConvertResponse
{
[JsonPropertyName("document")]
public DoclingDocument? Document { get; set; }
[JsonPropertyName("status")]
public string Status { get; set; } = string.Empty;
[JsonPropertyName("errors")]
public List<string>? Errors { get; set; }
}
public class DoclingDocument
{
[JsonPropertyName("md_content")]
public string? MarkdownContent { get; set; }
[JsonPropertyName("json_content")]
public object? JsonContent { get; set; }
[JsonPropertyName("text_content")]
public string? TextContent { get; set; }
[JsonPropertyName("metadata")]
public DoclingMetadata? Metadata { get; set; }
}
public class DoclingMetadata
{
[JsonPropertyName("filename")]
public string? Filename { get; set; }
[JsonPropertyName("page_count")]
public int? PageCount { get; set; }
[JsonPropertyName("file_type")]
public string? FileType { get; set; }
}
/// <summary>
/// Our internal document model
/// </summary>
public class ProcessedDocument
{
public string DocumentId { get; set; } = Guid.NewGuid().ToString();
public string FileName { get; set; } = string.Empty;
public string OriginalFormat { get; set; } = string.Empty;
public string MarkdownContent { get; set; } = string.Empty;
public string? TextContent { get; set; }
public DateTime ProcessedDate { get; set; } = DateTime.UtcNow;
public string[] Categories { get; set; } = Array.Empty<string>();
public int? PageCount { get; set; }
public string ContentHash { get; set; } = string.Empty;
}
}
using Microsoft.Extensions.Logging;
using Microsoft.Extensions.Options;
using Mostlylucid.BlogLLM.Core.Models;
using System.Net.Http.Json;
using System.Security.Cryptography;
using System.Text;
using System.Text.Json;
namespace Mostlylucid.BlogLLM.Core.Services
{
public class DoclingClientOptions
{
public string BaseUrl { get; set; } = "http://localhost:5001";
public int TimeoutSeconds { get; set; } = 300; // 5 minutes for large documents
public bool EnableOcr { get; set; } = true;
public string TableMode { get; set; } = "accurate";
}
public class DoclingClient : IDisposable
{
private readonly HttpClient _httpClient;
private readonly ILogger<DoclingClient> _logger;
private readonly DoclingClientOptions _options;
public DoclingClient(
HttpClient httpClient,
ILogger<DoclingClient> logger,
IOptions<DoclingClientOptions> options)
{
_httpClient = httpClient;
_logger = logger;
_options = options.Value;
_httpClient.BaseAddress = new Uri(_options.BaseUrl);
_httpClient.Timeout = TimeSpan.FromSeconds(_options.TimeoutSeconds);
}
/// <summary>
/// Convert a document from a URL
/// </summary>
public async Task<ProcessedDocument?> ConvertFromUrlAsync(
string url,
string[]? categories = null,
CancellationToken cancellationToken = default)
{
_logger.LogInformation("Converting document from URL: {Url}", url);
var request = new DoclingConvertRequest
{
Sources = new List<DoclingSource>
{
new() { Kind = "http", Url = url }
},
Options = new DoclingOptions
{
ToFormats = new List<string> { "md", "text" },
Ocr = _options.EnableOcr,
TableMode = _options.TableMode
}
};
return await SendConversionRequestAsync(request, categories, cancellationToken);
}
/// <summary>
/// Convert a local file
/// </summary>
public async Task<ProcessedDocument?> ConvertFileAsync(
string filePath,
string[]? categories = null,
CancellationToken cancellationToken = default)
{
if (!File.Exists(filePath))
{
_logger.LogError("File not found: {FilePath}", filePath);
return null;
}
_logger.LogInformation("Converting local file: {FilePath}", filePath);
// Read file and convert to base64
var fileBytes = await File.ReadAllBytesAsync(filePath, cancellationToken);
var base64Content = Convert.ToBase64String(fileBytes);
var fileName = Path.GetFileName(filePath);
var request = new DoclingConvertRequest
{
Sources = new List<DoclingSource>
{
new()
{
Kind = "base64",
Base64Content = base64Content,
Filename = fileName
}
},
Options = new DoclingOptions
{
ToFormats = new List<string> { "md", "text" },
Ocr = _options.EnableOcr,
TableMode = _options.TableMode
}
};
return await SendConversionRequestAsync(request, categories, cancellationToken);
}
/// <summary>
/// Convert a file using multipart form upload (more efficient for large files)
/// </summary>
public async Task<ProcessedDocument?> ConvertFileUploadAsync(
string filePath,
string[]? categories = null,
CancellationToken cancellationToken = default)
{
if (!File.Exists(filePath))
{
_logger.LogError("File not found: {FilePath}", filePath);
return null;
}
_logger.LogInformation("Uploading and converting file: {FilePath}", filePath);
try
{
using var fileStream = File.OpenRead(filePath);
using var content = new MultipartFormDataContent();
using var streamContent = new StreamContent(fileStream);
var fileName = Path.GetFileName(filePath);
content.Add(streamContent, "files", fileName);
var response = await _httpClient.PostAsync(
"/v1/convert/file",
content,
cancellationToken);
if (!response.IsSuccessStatusCode)
{
var errorContent = await response.Content.ReadAsStringAsync(cancellationToken);
_logger.LogError("Docling conversion failed: {StatusCode} - {Error}",
response.StatusCode, errorContent);
return null;
}
var result = await response.Content.ReadFromJsonAsync<DoclingConvertResponse>(
cancellationToken: cancellationToken);
return MapToProcessedDocument(result, fileName, categories);
}
catch (Exception ex)
{
_logger.LogError(ex, "Error uploading file to Docling: {FilePath}", filePath);
return null;
}
}
private async Task<ProcessedDocument?> SendConversionRequestAsync(
DoclingConvertRequest request,
string[]? categories,
CancellationToken cancellationToken)
{
try
{
var response = await _httpClient.PostAsJsonAsync(
"/v1/convert/source",
request,
cancellationToken);
if (!response.IsSuccessStatusCode)
{
var errorContent = await response.Content.ReadAsStringAsync(cancellationToken);
_logger.LogError("Docling conversion failed: {StatusCode} - {Error}",
response.StatusCode, errorContent);
return null;
}
var result = await response.Content.ReadFromJsonAsync<DoclingConvertResponse>(
cancellationToken: cancellationToken);
var filename = request.Sources.FirstOrDefault()?.Filename
?? request.Sources.FirstOrDefault()?.Url
?? "unknown";
return MapToProcessedDocument(result, filename, categories);
}
catch (HttpRequestException ex)
{
_logger.LogError(ex, "HTTP error calling Docling API");
return null;
}
catch (TaskCanceledException ex)
{
_logger.LogError(ex, "Docling conversion timed out");
return null;
}
catch (Exception ex)
{
_logger.LogError(ex, "Unexpected error calling Docling API");
return null;
}
}
private ProcessedDocument? MapToProcessedDocument(
DoclingConvertResponse? response,
string filename,
string[]? categories)
{
if (response?.Document == null)
{
_logger.LogWarning("Docling returned empty document");
return null;
}
var markdownContent = response.Document.MarkdownContent ?? string.Empty;
return new ProcessedDocument
{
DocumentId = Guid.NewGuid().ToString(),
FileName = Path.GetFileName(filename),
OriginalFormat = Path.GetExtension(filename).TrimStart('.').ToLower(),
MarkdownContent = markdownContent,
TextContent = response.Document.TextContent,
ProcessedDate = DateTime.UtcNow,
Categories = categories ?? Array.Empty<string>(),
PageCount = response.Document.Metadata?.PageCount,
ContentHash = ComputeHash(markdownContent)
};
}
private static string ComputeHash(string content)
{
using var sha256 = SHA256.Create();
var bytes = sha256.ComputeHash(Encoding.UTF8.GetBytes(content));
return Convert.ToBase64String(bytes);
}
public void Dispose()
{
_httpClient?.Dispose();
}
}
}
Ora integriamo Docling con la nostra pipeline RAG esistente:
using Microsoft.Extensions.Logging;
using Mostlylucid.BlogLLM.Core.Models;
namespace Mostlylucid.BlogLLM.Core.Services
{
public class DocumentIngestionService
{
private readonly ILogger<DocumentIngestionService> _logger;
private readonly DoclingClient _doclingClient;
private readonly MarkdownParserService _markdownParser;
private readonly ChunkingService _chunker;
private readonly BatchEmbeddingService _embedder;
private readonly QdrantVectorStore _vectorStore;
public DocumentIngestionService(
ILogger<DocumentIngestionService> logger,
DoclingClient doclingClient,
MarkdownParserService markdownParser,
ChunkingService chunker,
BatchEmbeddingService embedder,
QdrantVectorStore vectorStore)
{
_logger = logger;
_doclingClient = doclingClient;
_markdownParser = markdownParser;
_chunker = chunker;
_embedder = embedder;
_vectorStore = vectorStore;
}
/// <summary>
/// Process a document file and add to vector store
/// </summary>
public async Task<bool> ProcessDocumentAsync(
string filePath,
string[]? categories = null,
CancellationToken cancellationToken = default)
{
try
{
_logger.LogInformation("Processing document: {FilePath}", filePath);
// Step 1: Convert with Docling
var document = await _doclingClient.ConvertFileUploadAsync(
filePath, categories, cancellationToken);
if (document == null)
{
_logger.LogError("Failed to convert document: {FilePath}", filePath);
return false;
}
return await ProcessConvertedDocumentAsync(document, cancellationToken);
}
catch (Exception ex)
{
_logger.LogError(ex, "Error processing document {FilePath}", filePath);
return false;
}
}
/// <summary>
/// Process a document from URL and add to vector store
/// </summary>
public async Task<bool> ProcessDocumentFromUrlAsync(
string url,
string[]? categories = null,
CancellationToken cancellationToken = default)
{
try
{
_logger.LogInformation("Processing document from URL: {Url}", url);
// Step 1: Convert with Docling
var document = await _doclingClient.ConvertFromUrlAsync(
url, categories, cancellationToken);
if (document == null)
{
_logger.LogError("Failed to convert document from URL: {Url}", url);
return false;
}
return await ProcessConvertedDocumentAsync(document, cancellationToken);
}
catch (Exception ex)
{
_logger.LogError(ex, "Error processing document from URL {Url}", url);
return false;
}
}
private async Task<bool> ProcessConvertedDocumentAsync(
ProcessedDocument document,
CancellationToken cancellationToken)
{
_logger.LogInformation("Document converted: {FileName} ({PageCount} pages)",
document.FileName, document.PageCount ?? 0);
// Step 2: Parse the markdown content
var post = _markdownParser.ParseMarkdownFromContent(
document.MarkdownContent,
document.FileName);
post.Categories = document.Categories;
// Step 3: Chunk the content
var chunks = _chunker.ChunkBlogPost(post);
_logger.LogInformation("Created {ChunkCount} chunks from {FileName}",
chunks.Count, document.FileName);
if (chunks.Count == 0)
{
_logger.LogWarning("No chunks created for document: {FileName}", document.FileName);
return false;
}
// Step 4: Generate embeddings
var progress = new Progress<int>(processed =>
{
_logger.LogDebug("Embedded {Processed}/{Total} chunks",
processed, chunks.Count);
});
await _embedder.GenerateEmbeddingsAsync(chunks, progress, cancellationToken);
// Step 5: Store in vector database
await _vectorStore.UpsertChunksAsync(chunks);
_logger.LogInformation("Successfully processed document: {FileName} ({ChunkCount} chunks)",
document.FileName, chunks.Count);
return true;
}
/// <summary>
/// Batch process multiple documents
/// </summary>
public async Task<(int success, int failed)> ProcessDocumentBatchAsync(
IEnumerable<string> filePaths,
string[]? categories = null,
CancellationToken cancellationToken = default)
{
int success = 0;
int failed = 0;
foreach (var filePath in filePaths)
{
if (cancellationToken.IsCancellationRequested)
break;
var result = await ProcessDocumentAsync(filePath, categories, cancellationToken);
if (result)
success++;
else
failed++;
}
_logger.LogInformation("Batch processing complete: {Success} succeeded, {Failed} failed",
success, failed);
return (success, failed);
}
}
}
using Microsoft.Extensions.DependencyInjection;
using Mostlylucid.BlogLLM.Core.Services;
public static class ServiceCollectionExtensions
{
public static IServiceCollection AddDoclingServices(
this IServiceCollection services,
Action<DoclingClientOptions>? configureOptions = null)
{
// Configure options
if (configureOptions != null)
{
services.Configure(configureOptions);
}
else
{
services.Configure<DoclingClientOptions>(options =>
{
options.BaseUrl = "http://localhost:5001";
options.TimeoutSeconds = 300;
options.EnableOcr = true;
});
}
// Register HttpClient with configuration
services.AddHttpClient<DoclingClient>((serviceProvider, client) =>
{
var options = serviceProvider
.GetRequiredService<IOptions<DoclingClientOptions>>().Value;
client.BaseAddress = new Uri(options.BaseUrl);
client.Timeout = TimeSpan.FromSeconds(options.TimeoutSeconds);
});
// Register services
services.AddScoped<DocumentIngestionService>();
return services;
}
}
Aggiungi al tuo appsettings.json:
{
"Docling": {
"BaseUrl": "http://localhost:5001",
"TimeoutSeconds": 300,
"EnableOcr": true,
"TableMode": "accurate"
}
}
// PDF Processing
await documentIngestionService.ProcessDocumentAsync(
"C:\\documents\\client_contract.pdf",
new[] { "legal", "contracts", "client-agreements" });
// DOCX Processing
await documentIngestionService.ProcessDocumentAsync(
"C:\\documents\\motion_to_dismiss.docx",
new[] { "legal", "briefs", "motions" });
// URL Processing (great for public documents)
await documentIngestionService.ProcessDocumentFromUrlAsync(
"https://arxiv.org/pdf/2501.17887",
new[] { "research", "ai", "docling" });
// Batch Processing
var files = Directory.GetFiles("C:\\documents\\legal", "*.pdf");
var (success, failed) = await documentIngestionService.ProcessDocumentBatchAsync(
files,
new[] { "legal", "batch-import" });
Console.WriteLine($"Processed {success} files, {failed} failures");
public class DocumentWatcherService : BackgroundService
{
private readonly IServiceProvider _serviceProvider;
private readonly ILogger<DocumentWatcherService> _logger;
private FileSystemWatcher? _watcher;
private readonly string _watchPath;
public DocumentWatcherService(
IServiceProvider serviceProvider,
ILogger<DocumentWatcherService> logger,
IConfiguration configuration)
{
_serviceProvider = serviceProvider;
_logger = logger;
_watchPath = configuration["DocumentWatch:Path"] ?? "C:\\documents\\incoming";
}
protected override Task ExecuteAsync(CancellationToken stoppingToken)
{
if (!Directory.Exists(_watchPath))
{
Directory.CreateDirectory(_watchPath);
}
_watcher = new FileSystemWatcher(_watchPath)
{
NotifyFilter = NotifyFilters.FileName | NotifyFilters.LastWrite,
IncludeSubdirectories = false
};
// Watch for common document types
_watcher.Filters.Add("*.pdf");
_watcher.Filters.Add("*.docx");
_watcher.Filters.Add("*.doc");
_watcher.Filters.Add("*.pptx");
_watcher.Created += OnFileCreated;
_watcher.EnableRaisingEvents = true;
_logger.LogInformation("Watching for documents in: {Path}", _watchPath);
return Task.CompletedTask;
}
private async void OnFileCreated(object sender, FileSystemEventArgs e)
{
_logger.LogInformation("New document detected: {FileName}", e.Name);
// Wait for file to be fully written
await Task.Delay(1000);
using var scope = _serviceProvider.CreateScope();
var ingestionService = scope.ServiceProvider
.GetRequiredService<DocumentIngestionService>();
await ingestionService.ProcessDocumentAsync(e.FullPath);
}
public override void Dispose()
{
_watcher?.Dispose();
base.Dispose();
}
}
| Tipo di documento | Dimensione tipica | Lavorazione di aggancio | Tempo di inserimento |
|---|---|---|---|
| PDF semplice (1-5 pagine) | 100KB-500KB | 2-5 secondi | 0,5-1 secondo |
| PDF complesso (20+ pagine) | 1-5MB | 10-30 secondi | 2-5 secondi |
| PDF scansionato con OCR | 1-10MB | 30-120 secondi | 2-5 secondi |
| Documento DOCX | 50KB-500KB | 1-3 secondi | 0,3-0,8 secondi |
| Presentazione PPTX | 1-20MB | 5-30 secondi | 1-3 secondi |
docling-serve-cu126 oppure docling-serve-cu128)fast modalità tabella quando l'accuratezza della tabella non è criticaAbbiamo integrato con successo Docling nel nostro sistema GPT dell'avvocato, consentendo:
Questo amplia la nostra base di conoscenze oltre i post del blog per includere documenti legali, contratti, documenti di ricerca e altri materiali importanti.
© 2026 Scott Galloway — Unlicense — All content and source code on this site is free to use, copy, modify, and sell.