mirror of
https://github.com/microsoft/agent-framework.git
synced 2026-06-16 21:04:09 +08:00
* .NET: Add a TODO AIContextProvider (#5233) * Add a TODO AIContextProvider * Add unit tests * Address PR comments * Address PR comments * Fix test after removing one tool * .NET: Add a ModeProvider for managing agent modes (#5247) * Add a ModeProvider for managing agent modes * Fix typo * Fix typo * Fix typo * Address PR comments * .NET: Add sample to show how to build a harness (#5268) * Add sample to show how to build a harness * Improve sample * Sample max output tokens and model * Fix encoding * Fix model name in readme * Address PR comments * .NET: Add context window size compaction strategy for harness (#5304) * Add context window size compaction strategy for harness * Apply suggestions from code review Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * Address PR comments --------- Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * .NET: Add a file memory provider (#5315) * Add a file memory provider * Address PR comments * Fix review comments. * Add additional unit tests * Addressing PR comments. * .NET: Harness: Improve prompts and add FileSystem store (#5365) * Harness: Improve prompts and add FileSystem store * Address PR comments * .NET: Harness: Improve path validation (#5404) * Harness: Improve path validation * Address PR comments * .NET: Add always approve helpers, improve sample and fix bug (#5451) * Add always approve helpers, improve sample and fix bug * Address PR comments * .NET: Make Todo, Mode and FileMemory providers more configurable (#5477) * Make Todo, Mode and FileMemory providers more configurable * Address PR comments. * .NET: Add subagents provider and sample (#5518) * Add subagents provider and sample * Addressing PR comments. * .NET: Harness filememory index plus instructions consistency (#5540) * Add FileMemoryProvider index and improve instruction consistency * Address PR comments. * Address PR comments * Address PR comments. * Apply suggestion from @rogerbarreto Co-authored-by: Roger Barreto <19890735+rogerbarreto@users.noreply.github.com> --------- Co-authored-by: Roger Barreto <19890735+rogerbarreto@users.noreply.github.com> * .NET: Refactor harness console to be more extensible and easy to understand with better UX (#5573) * Refactor harness console to be more extensible and easy to understand with better UX. * Fix formatting issues. * Allow multiple clarifications in one response * Address PR comments * .NET: Add FileAccessProvdider and concurrency fix for FileMemoryProvider (#5583) * Add FileAccessProvdider and concurrency fix for FileMemoryProvider * Address PR comments --------- Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Co-authored-by: Roger Barreto <19890735+rogerbarreto@users.noreply.github.com>
149 lines
6.5 KiB
C#
149 lines
6.5 KiB
C#
// Copyright (c) Microsoft. All rights reserved.
|
|
|
|
using System;
|
|
using System.Diagnostics.CodeAnalysis;
|
|
using System.Threading;
|
|
using System.Threading.Tasks;
|
|
using Microsoft.Extensions.Logging;
|
|
using Microsoft.Shared.DiagnosticIds;
|
|
using Microsoft.Shared.Diagnostics;
|
|
|
|
namespace Microsoft.Agents.AI.Compaction;
|
|
|
|
/// <summary>
|
|
/// A compaction strategy that derives token thresholds from a model's context window size
|
|
/// and maximum output tokens, applying a two-phase compaction pipeline:
|
|
/// <list type="number">
|
|
/// <item><description><b>Tool result eviction</b> (<see cref="ToolResultCompactionStrategy"/>) — collapses old tool call groups
|
|
/// into concise summaries when the token count exceeds the <see cref="ToolEvictionThreshold"/>.</description></item>
|
|
/// <item><description><b>Truncation</b> (<see cref="TruncationCompactionStrategy"/>) — removes the oldest non-system message groups
|
|
/// when the token count exceeds the <see cref="TruncationThreshold"/>.</description></item>
|
|
/// </list>
|
|
/// </summary>
|
|
/// <remarks>
|
|
/// <para>
|
|
/// The <b>input budget</b> is defined as <c>maxContextWindowTokens - maxOutputTokens</c>, representing
|
|
/// the maximum number of tokens available for the conversation input (including system messages, tools, and history).
|
|
/// </para>
|
|
/// <para>
|
|
/// This strategy is a convenience wrapper around <see cref="PipelineCompactionStrategy"/> that automates
|
|
/// threshold calculation from model specifications.
|
|
/// </para>
|
|
/// </remarks>
|
|
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
|
|
public sealed class ContextWindowCompactionStrategy : CompactionStrategy
|
|
{
|
|
/// <summary>
|
|
/// The default fraction of the input budget at which tool result eviction triggers.
|
|
/// </summary>
|
|
public const double DefaultToolEvictionThreshold = 0.5;
|
|
|
|
/// <summary>
|
|
/// The default fraction of the input budget at which truncation triggers.
|
|
/// </summary>
|
|
public const double DefaultTruncationThreshold = 0.8;
|
|
|
|
private readonly PipelineCompactionStrategy _pipeline;
|
|
|
|
/// <summary>
|
|
/// Initializes a new instance of the <see cref="ContextWindowCompactionStrategy"/> class.
|
|
/// </summary>
|
|
/// <param name="maxContextWindowTokens">
|
|
/// The maximum number of tokens the model's context window supports (e.g., 1,050,000 for gpt-5.4).
|
|
/// </param>
|
|
/// <param name="maxOutputTokens">
|
|
/// The maximum number of output tokens the model can generate per response (e.g., 128,000 for gpt-5.4).
|
|
/// </param>
|
|
/// <param name="toolEvictionThreshold">
|
|
/// The fraction of the input budget (0.0, 1.0] at which tool result eviction triggers.
|
|
/// Defaults to <see cref="DefaultToolEvictionThreshold"/> (0.5).
|
|
/// </param>
|
|
/// <param name="truncationThreshold">
|
|
/// The fraction of the input budget (0.0, 1.0] at which truncation triggers.
|
|
/// Defaults to <see cref="DefaultTruncationThreshold"/> (0.8).
|
|
/// Must be greater than or equal to <paramref name="toolEvictionThreshold"/>.
|
|
/// </param>
|
|
/// <exception cref="ArgumentOutOfRangeException">
|
|
/// <paramref name="maxContextWindowTokens"/> is not positive, or
|
|
/// <paramref name="maxOutputTokens"/> is negative or greater than or equal to <paramref name="maxContextWindowTokens"/>, or
|
|
/// <paramref name="toolEvictionThreshold"/> or <paramref name="truncationThreshold"/> is not in (0.0, 1.0], or
|
|
/// <paramref name="truncationThreshold"/> is less than <paramref name="toolEvictionThreshold"/>.
|
|
/// </exception>
|
|
public ContextWindowCompactionStrategy(
|
|
int maxContextWindowTokens,
|
|
int maxOutputTokens,
|
|
double toolEvictionThreshold = DefaultToolEvictionThreshold,
|
|
double truncationThreshold = DefaultTruncationThreshold)
|
|
: base(CompactionTriggers.Always)
|
|
{
|
|
Throw.IfLessThanOrEqual(maxContextWindowTokens, 0);
|
|
Throw.IfLessThan(maxOutputTokens, 0);
|
|
Throw.IfGreaterThanOrEqual(maxOutputTokens, maxContextWindowTokens);
|
|
|
|
ValidateThreshold(toolEvictionThreshold, nameof(toolEvictionThreshold));
|
|
ValidateThreshold(truncationThreshold, nameof(truncationThreshold));
|
|
|
|
if (truncationThreshold < toolEvictionThreshold)
|
|
{
|
|
throw new ArgumentOutOfRangeException(nameof(truncationThreshold), truncationThreshold,
|
|
$"Truncation threshold ({truncationThreshold}) must be greater than or equal to tool eviction threshold ({toolEvictionThreshold}).");
|
|
}
|
|
|
|
this.MaxContextWindowTokens = maxContextWindowTokens;
|
|
this.MaxOutputTokens = maxOutputTokens;
|
|
this.InputBudgetTokens = maxContextWindowTokens - maxOutputTokens;
|
|
this.ToolEvictionThreshold = toolEvictionThreshold;
|
|
this.TruncationThreshold = truncationThreshold;
|
|
|
|
int toolEvictionTokens = (int)(this.InputBudgetTokens * toolEvictionThreshold);
|
|
int truncationTokens = (int)(this.InputBudgetTokens * truncationThreshold);
|
|
|
|
this._pipeline = new PipelineCompactionStrategy(
|
|
new ToolResultCompactionStrategy(
|
|
trigger: CompactionTriggers.TokensExceed(toolEvictionTokens),
|
|
minimumPreservedGroups: 2),
|
|
new TruncationCompactionStrategy(
|
|
trigger: CompactionTriggers.TokensExceed(truncationTokens),
|
|
minimumPreservedGroups: 2));
|
|
}
|
|
|
|
/// <summary>
|
|
/// Gets the maximum context window size in tokens.
|
|
/// </summary>
|
|
public int MaxContextWindowTokens { get; }
|
|
|
|
/// <summary>
|
|
/// Gets the maximum output tokens per response.
|
|
/// </summary>
|
|
public int MaxOutputTokens { get; }
|
|
|
|
/// <summary>
|
|
/// Gets the computed input budget in tokens (<see cref="MaxContextWindowTokens"/> minus <see cref="MaxOutputTokens"/>).
|
|
/// </summary>
|
|
public int InputBudgetTokens { get; }
|
|
|
|
/// <summary>
|
|
/// Gets the fraction of the input budget at which tool result eviction triggers.
|
|
/// </summary>
|
|
public double ToolEvictionThreshold { get; }
|
|
|
|
/// <summary>
|
|
/// Gets the fraction of the input budget at which truncation triggers.
|
|
/// </summary>
|
|
public double TruncationThreshold { get; }
|
|
|
|
/// <inheritdoc/>
|
|
protected override async ValueTask<bool> CompactCoreAsync(CompactionMessageIndex index, ILogger logger, CancellationToken cancellationToken)
|
|
{
|
|
return await this._pipeline.CompactAsync(index, logger, cancellationToken).ConfigureAwait(false);
|
|
}
|
|
|
|
private static void ValidateThreshold(double value, string paramName)
|
|
{
|
|
if (value is <= 0.0 or > 1.0)
|
|
{
|
|
throw new ArgumentOutOfRangeException(paramName, value, "Threshold must be in the range (0.0, 1.0].");
|
|
}
|
|
}
|
|
}
|