171 lines
6.9 KiB
C#
171 lines
6.9 KiB
C#
// Copyright (c) Microsoft. All rights reserved.
|
||
|
||
using System.Net;
|
||
using Microsoft.Extensions.VectorData;
|
||
using Microsoft.SemanticKernel;
|
||
using Microsoft.SemanticKernel.ChatCompletion;
|
||
using UglyToad.PdfPig;
|
||
using UglyToad.PdfPig.Content;
|
||
using UglyToad.PdfPig.DocumentLayoutAnalysis.PageSegmenter;
|
||
|
||
namespace VectorStoreRAG;
|
||
|
||
/// <summary>
|
||
/// Class that loads text from a PDF file into a vector store.
|
||
/// </summary>
|
||
/// <typeparam name="TKey">The type of the data model key.</typeparam>
|
||
/// <param name="uniqueKeyGenerator">A function to generate unique keys with.</param>
|
||
/// <param name="vectorStoreRecordCollection">The collection to load the data into.</param>
|
||
/// <param name="chatCompletionService">The chat completion service to use for generating text from images.</param>
|
||
internal sealed class DataLoader<TKey>(
|
||
UniqueKeyGenerator<TKey> uniqueKeyGenerator,
|
||
VectorStoreCollection<TKey, TextSnippet<TKey>> vectorStoreRecordCollection,
|
||
IChatCompletionService chatCompletionService) : IDataLoader where TKey : notnull
|
||
{
|
||
/// <inheritdoc/>
|
||
public async Task LoadPdf(string pdfPath, int batchSize, int betweenBatchDelayInMs, CancellationToken cancellationToken)
|
||
{
|
||
// Create the collection if it doesn't exist.
|
||
await vectorStoreRecordCollection.EnsureCollectionExistsAsync(cancellationToken).ConfigureAwait(false);
|
||
|
||
// Load the text and images from the PDF file and split them into batches.
|
||
var sections = LoadTextAndImages(pdfPath, cancellationToken);
|
||
var batches = sections.Chunk(batchSize);
|
||
|
||
// Process each batch of content items.
|
||
foreach (var batch in batches)
|
||
{
|
||
// Convert any images to text.
|
||
var textContentTasks = batch.Select(async content =>
|
||
{
|
||
if (content.Text != null)
|
||
{
|
||
return content;
|
||
}
|
||
|
||
var textFromImage = await ConvertImageToTextWithRetryAsync(
|
||
chatCompletionService,
|
||
content.Image!.Value,
|
||
cancellationToken).ConfigureAwait(false);
|
||
return new RawContent { Text = textFromImage, PageNumber = content.PageNumber };
|
||
});
|
||
var textContent = await Task.WhenAll(textContentTasks).ConfigureAwait(false);
|
||
|
||
// Map each paragraph to a TextSnippet.
|
||
var records = textContent.Select(content => new TextSnippet<TKey>
|
||
{
|
||
Key = uniqueKeyGenerator.GenerateKey(),
|
||
// The vector store will automatically generate the embedding for this text.
|
||
// See the TextEmbedding field on the TextSnippet class.
|
||
Text = content.Text,
|
||
ReferenceDescription = $"{new FileInfo(pdfPath).Name}#page={content.PageNumber}",
|
||
ReferenceLink = $"{new Uri(new FileInfo(pdfPath).FullName).AbsoluteUri}#page={content.PageNumber}",
|
||
});
|
||
|
||
// Upsert the records into the vector store.
|
||
await vectorStoreRecordCollection.UpsertAsync(records, cancellationToken: cancellationToken).ConfigureAwait(false);
|
||
|
||
await Task.Delay(betweenBatchDelayInMs, cancellationToken).ConfigureAwait(false);
|
||
}
|
||
}
|
||
|
||
/// <summary>
|
||
/// Read the text and images from each page in the provided PDF file.
|
||
/// </summary>
|
||
/// <param name="pdfPath">The pdf file to read the text and images from.</param>
|
||
/// <param name="cancellationToken">The <see cref="CancellationToken"/> to monitor for cancellation requests.</param>
|
||
/// <returns>The text and images from the pdf file, plus the page number that each is on.</returns>
|
||
private static IEnumerable<RawContent> LoadTextAndImages(string pdfPath, CancellationToken cancellationToken)
|
||
{
|
||
using (PdfDocument document = PdfDocument.Open(pdfPath))
|
||
{
|
||
foreach (Page page in document.GetPages())
|
||
{
|
||
if (cancellationToken.IsCancellationRequested)
|
||
{
|
||
break;
|
||
}
|
||
|
||
foreach (var image in page.GetImages())
|
||
{
|
||
if (image.TryGetPng(out var png))
|
||
{
|
||
yield return new RawContent { Image = png, PageNumber = page.Number };
|
||
}
|
||
else
|
||
{
|
||
Console.WriteLine($"Unsupported image format on page {page.Number}");
|
||
}
|
||
}
|
||
|
||
var blocks = DefaultPageSegmenter.Instance.GetBlocks(page.GetWords());
|
||
foreach (var block in blocks)
|
||
{
|
||
if (cancellationToken.IsCancellationRequested)
|
||
{
|
||
break;
|
||
}
|
||
|
||
yield return new RawContent { Text = block.Text, PageNumber = page.Number };
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
/// <summary>
|
||
/// Add a simple retry mechanism to image to text.
|
||
/// </summary>
|
||
/// <param name="chatCompletionService">The chat completion service to use for generating text from images.</param>
|
||
/// <param name="imageBytes">The image to generate the text for.</param>
|
||
/// <param name="cancellationToken">The <see cref="CancellationToken"/> to monitor for cancellation requests.</param>
|
||
/// <returns>The generated text.</returns>
|
||
private static async Task<string> ConvertImageToTextWithRetryAsync(
|
||
IChatCompletionService chatCompletionService,
|
||
ReadOnlyMemory<byte> imageBytes,
|
||
CancellationToken cancellationToken)
|
||
{
|
||
var tries = 0;
|
||
|
||
while (true)
|
||
{
|
||
try
|
||
{
|
||
var chatHistory = new ChatHistory();
|
||
chatHistory.AddUserMessage([
|
||
new TextContent("What’s in this image?"),
|
||
new ImageContent(imageBytes, "image/png"),
|
||
]);
|
||
var result = await chatCompletionService.GetChatMessageContentsAsync(chatHistory, cancellationToken: cancellationToken).ConfigureAwait(false);
|
||
return string.Join("\n", result.Select(x => x.Content));
|
||
}
|
||
catch (HttpOperationException ex) when (ex.StatusCode == HttpStatusCode.TooManyRequests)
|
||
{
|
||
tries++;
|
||
|
||
if (tries < 3)
|
||
{
|
||
Console.WriteLine($"Failed to generate text from image. Error: {ex}");
|
||
Console.WriteLine("Retrying text to image conversion...");
|
||
await Task.Delay(10_000, cancellationToken).ConfigureAwait(false);
|
||
}
|
||
else
|
||
{
|
||
throw;
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
/// <summary>
|
||
/// Private model for returning the content items from a PDF file.
|
||
/// </summary>
|
||
private sealed class RawContent
|
||
{
|
||
public string? Text { get; init; }
|
||
|
||
public ReadOnlyMemory<byte>? Image { get; init; }
|
||
|
||
public int PageNumber { get; init; }
|
||
}
|
||
}
|