AI-Studio/app/MindWork AI Studio/Tools/ToolCallingSystem/ToolCallingImplementations/WebSearch/WebSearchResultRetrievalService.cs

228 lines
10 KiB
C#

using AIStudio.Tools.Web;
namespace AIStudio.Tools.ToolCallingSystem.ToolCallingImplementations.WebSearch;
internal sealed class WebSearchResultRetrievalService(WebPageRetrievalService webPageRetrievalService)
{
private const int MAX_PARALLEL_RETRIEVALS = 4;
/// <summary>
/// How much of a search service's snippet a hit without a readable page may return.
/// </summary>
/// <remarks>
/// A snippet is a sentence or two by design, so this limit is never reached by a service
/// behaving as documented. It exists so that one that does not cannot smuggle text past the
/// content budget, which is a setting the user made and which snippets do not draw from.
/// </remarks>
private const int MAX_SNIPPET_CHARACTERS = 1000;
public async Task<WebSearchPageRetrievalResult> RetrieveAsync(
IReadOnlyList<SearchCandidate> candidates,
int pageTimeoutSeconds,
int allPagesRetrievalTimeoutSeconds,
int maxTotalContentCharacters,
int minContentCharactersPerResult,
CancellationToken token)
{
var counters = new RetrievalCounters();
using var retrievalTimeoutCts = CancellationTokenSource.CreateLinkedTokenSource(token);
retrievalTimeoutCts.CancelAfter(TimeSpan.FromSeconds(allPagesRetrievalTimeoutSeconds));
using var retrievalSemaphore = new SemaphoreSlim(MAX_PARALLEL_RETRIEVALS);
//
// Started in a loop rather than through a Select: a lambda would capture the semaphore and
// the timeout source, and a captured disposable outliving its scope is exactly what one
// cannot see from the call site. Handing them over as arguments keeps that impossible.
//
var retrievalTasks = new List<Task<WebSearchPageResult>>(candidates.Count);
foreach (var candidate in candidates)
retrievalTasks.Add(this.RetrieveCandidateAsync(candidate, pageTimeoutSeconds, retrievalSemaphore, retrievalTimeoutCts, counters, token));
var retrievedPages = await Task.WhenAll(retrievalTasks);
token.ThrowIfCancellationRequested();
var mergedResults = MergeDuplicates(retrievedPages);
ApplySnippets(mergedResults);
ApplyContentBudget(mergedResults, maxTotalContentCharacters, minContentCharactersPerResult);
var statistics = new WebSearchPageRetrievalStatistics(
counters.Attempted,
counters.Blocked,
counters.PageTimedOut,
counters.Failed,
counters.EmptyContent);
return new WebSearchPageRetrievalResult(mergedResults, counters.RetrievalTimedOut == 1, statistics);
}
/// <summary>
/// Retrieves one search result page, counting how it went.
/// </summary>
/// <remarks>
/// The semaphore and the timeout source belong to the caller, which disposes them once every
/// retrieval has finished. Passing them in rather than capturing them keeps that ownership
/// visible: nothing here outlives the call that hands them over.<br/><br/>
/// Every candidate comes back, whether or not its page could be read. A hit the search
/// service found is worth reporting even without its content: the model can still name the
/// page as a place to look, and the snippet often answers the question by itself. Only a
/// candidate that has nothing left to say is dropped, and that is decided after merging.
/// </remarks>
private async Task<WebSearchPageResult> RetrieveCandidateAsync(
SearchCandidate candidate,
int pageTimeoutSeconds,
SemaphoreSlim retrievalSemaphore,
CancellationTokenSource retrievalTimeoutCts,
RetrievalCounters counters,
CancellationToken token)
{
var enteredSemaphore = false;
try
{
await retrievalSemaphore.WaitAsync(retrievalTimeoutCts.Token);
enteredSemaphore = true;
Interlocked.Increment(ref counters.Attempted);
var retrievedPage = await webPageRetrievalService.RetrieveAsync(
candidate.RetrievalUrl,
new WebPageRetrievalOptions
{
TimeoutSeconds = pageTimeoutSeconds,
PublicTargetsOnly = true,
},
retrievalTimeoutCts.Token);
if (string.IsNullOrWhiteSpace(retrievedPage.ExtractedPage.Markdown))
{
Interlocked.Increment(ref counters.EmptyContent);
return new(candidate, null, WebSearchPageRetrievalOutcome.NO_READABLE_CONTENT);
}
return new(candidate, retrievedPage, WebSearchPageRetrievalOutcome.RETRIEVED);
}
catch (OperationCanceledException) when (!token.IsCancellationRequested)
{
Interlocked.Exchange(ref counters.RetrievalTimedOut, 1);
return new(candidate, null, WebSearchPageRetrievalOutcome.RETRIEVAL_TIMED_OUT);
}
catch (WebPageAccessBlockedException)
{
Interlocked.Increment(ref counters.Blocked);
return new(candidate, null, WebSearchPageRetrievalOutcome.BLOCKED);
}
catch (TimeoutException)
{
Interlocked.Increment(ref counters.PageTimedOut);
return new(candidate, null, WebSearchPageRetrievalOutcome.PAGE_TIMED_OUT);
}
catch (InvalidOperationException)
{
Interlocked.Increment(ref counters.Failed);
return new(candidate, null, WebSearchPageRetrievalOutcome.FAILED);
}
finally
{
if (enteredSemaphore)
retrievalSemaphore.Release();
}
}
private static List<WebSearchPageResult> MergeDuplicates(IEnumerable<WebSearchPageResult> results) => results
.GroupBy(result => SearchCandidate.NormalizeUrl(result.CitationUrl), StringComparer.Ordinal)
.Select(group =>
{
//
// A page that was read carries its group even when a hit without one ranked better.
// The group is one page, and letting a snippet win would throw away the only thing
// the retrieval accomplished. Its rank and reported title still come from the best
// hit of the group, because that is what Merge takes from whichever ranked highest.
//
var rankedGroup = group.OrderByDescending(result => result.HasPageContent).ThenBy(result => result.Candidate.Rank).ToList();
var carrier = rankedGroup[0];
var metadata = carrier.Candidate.Clone();
foreach (var duplicate in rankedGroup.Skip(1))
metadata.Merge(duplicate.Candidate);
return new WebSearchPageResult(metadata, carrier.RetrievedPage, carrier.Outcome);
})
.Where(HasSomethingToReport)
.OrderBy(result => result.Candidate.Rank)
.ToList();
/// <summary>
/// Whether this hit still tells the model something once its page turned out to be
/// unreadable.
/// </summary>
/// <remarks>
/// Decided after merging, because a hit two services found may owe its title to one of them
/// and its snippet to the other. What remains here is a bare URL with no title and no
/// snippet, which costs tokens and says nothing, so it is dropped as it always was.
/// </remarks>
private static bool HasSomethingToReport(WebSearchPageResult result) =>
result.HasPageContent ||
!string.IsNullOrWhiteSpace(result.Candidate.Snippet) ||
!string.IsNullOrWhiteSpace(result.Candidate.Title);
/// <summary>
/// Puts the search service's snippet in place of the content of every page that could not
/// be read.
/// </summary>
private static void ApplySnippets(List<WebSearchPageResult> results)
{
foreach (var result in results)
{
if (result.HasPageContent)
continue;
var snippet = result.Candidate.Snippet;
result.ReturnedMarkdown = snippet.Length <= MAX_SNIPPET_CHARACTERS ? snippet : $"{snippet[..(MAX_SNIPPET_CHARACTERS - 1)].TrimEnd()}…";
}
}
/// <summary>
/// Shares the content budget between the pages that were read.
/// </summary>
/// <remarks>
/// Only they take part in it. The budget exists so that a few long pages do not crowd each
/// other out, and a hit returning a snippet has nothing to crowd with: reserving the
/// per-result minimum for it would let a page nobody could read shorten one somebody can.
/// The snippets are capped on their own instead.
/// </remarks>
private static void ApplyContentBudget(List<WebSearchPageResult> results, int maxTotalContentCharacters, int minContentCharactersPerResult)
{
var pageResults = results.Where(result => result.HasPageContent).ToList();
var remainingBudget = maxTotalContentCharacters;
for (var index = 0; index < pageResults.Count; index++)
{
var result = pageResults[index];
var originalMarkdown = result.RetrievedPage!.ExtractedPage.Markdown;
var remainingResults = pageResults.Count - index - 1;
var currentBudget = remainingBudget - minContentCharactersPerResult * remainingResults;
if (originalMarkdown.Length > currentBudget)
{
result.ReturnedMarkdown = MarkdownTruncator.Truncate(originalMarkdown, currentBudget);
result.ContentTruncated = true;
}
else
{
result.ReturnedMarkdown = originalMarkdown;
}
remainingBudget -= result.ReturnedMarkdown.Length;
}
}
/// <summary>
/// What became of the pages of one search, counted while they are fetched in parallel.
/// </summary>
/// <remarks>
/// Public fields rather than properties, because the retrievals count through Interlocked,
/// which needs a reference to the storage itself.<br/><br/>
/// These count retrievals, while the outcome on each result describes one hit. The two do
/// not have to agree: two hits leading to the same page are two retrievals and one result.
/// </remarks>
private sealed class RetrievalCounters
{
public int Attempted;
public int Blocked;
public int PageTimedOut;
public int Failed;
public int EmptyContent;
public int RetrievalTimedOut;
}
}