codex wip

This commit is contained in:
2026-07-06 18:36:39 +02:00
parent 2ada187ae0
commit 645189f21c
9 changed files with 829 additions and 10 deletions
+10 -1
View File
@@ -97,6 +97,7 @@ namespace AnotherReplayReader
// 外部注入:每次请求前调用获取最新配置
// 委托类型变更
public Func<AiRequestContext>? GetRequestContext { get; set; }
public Func<AiPromptSettings>? GetPromptSettings { get; set; }
// ---------- methods ----------
@@ -248,7 +249,7 @@ namespace AnotherReplayReader
FinishCurrentContent();
var requestContext = GetRequestContext();
var systemPrompt = AIAnalyze.GetSystemPrompt(replay, players);
var systemPrompt = AIAnalyze.GetSystemPrompt(replay, players, GetPromptSettings?.Invoke());
var userPrompt = AIAnalyze.BuildUserPrompt(replay.Mod, players, replayData, out var userPromptPrefix);
AppendLog($"让 AI 了解录像...",
userPromptPrefix
@@ -358,6 +359,14 @@ namespace AnotherReplayReader
requestContext,
OnChunk,
_linkedCts.Token);
var validationResult = AIAnalysisValidation.ValidateMachineReadableClaims(segmentResult.Response);
if (validationResult.HasIssues)
{
AppendLog(
$"第{segmentIndex}段机器可读声明检查",
AIAnalysisValidation.FormatIssues(validationResult.Issues),
false);
}
_lastSuccessfulSegment = lastState.CurrentSegment;
lastState = segmentResult.State;
UpdateTokenDisplay(segmentResult);
+52 -1
View File
@@ -32,6 +32,7 @@
<!-- 右侧:Provider 编辑 + 模型管理 -->
<Grid Grid.Column="1" Margin="10,0,0,0">
<Grid.RowDefinitions>
<RowDefinition Height="Auto"/>
<RowDefinition Height="Auto"/>
<RowDefinition Height="*"/>
</Grid.RowDefinitions>
@@ -69,8 +70,58 @@
</StackPanel>
</Grid>
<!-- Prompt 管理区域 -->
<GroupBox Header="提示词" Grid.Row="1" Margin="0,5,0,0">
<Grid Margin="5">
<Grid.RowDefinitions>
<RowDefinition Height="Auto"/>
<RowDefinition Height="Auto"/>
<RowDefinition Height="Auto"/>
</Grid.RowDefinitions>
<Grid.ColumnDefinitions>
<ColumnDefinition Width="100"/>
<ColumnDefinition Width="*"/>
</Grid.ColumnDefinitions>
<CheckBox x:Name="_useCustomPromptCheck"
Grid.Row="0" Grid.Column="1"
Content="完全使用自定义 System Prompt"
Margin="0,0,0,5"/>
<Label Grid.Row="1" Grid.Column="0" Content="System Prompt"/>
<TextBox x:Name="_customPromptBox"
Grid.Row="1" Grid.Column="1"
MinHeight="80"
MaxHeight="160"
AcceptsReturn="True"
TextWrapping="Wrap"
ScrollViewer.VerticalScrollBarVisibility="Auto"/>
<Grid Grid.Row="2" Grid.ColumnSpan="2" Margin="0,5,0,0">
<Grid.ColumnDefinitions>
<ColumnDefinition Width="100"/>
<ColumnDefinition Width="*"/>
<ColumnDefinition Width="Auto"/>
</Grid.ColumnDefinitions>
<Label Grid.Column="0" Content="补充规则"/>
<TextBox x:Name="_additionalRulesBox"
Grid.Column="1"
MinHeight="50"
MaxHeight="100"
AcceptsReturn="True"
TextWrapping="Wrap"
ScrollViewer.VerticalScrollBarVisibility="Auto"/>
<StackPanel Grid.Column="2" Margin="5,0,0,0">
<Button Content="应用提示词" Click="OnApplyPromptClick" Margin="2"/>
<Button Content="恢复默认" Click="OnResetPromptClick" Margin="2"/>
<Button Content="清空补充" Click="OnClearAdditionalRulesClick" Margin="2"/>
</StackPanel>
</Grid>
</Grid>
</GroupBox>
<!-- 模型管理区域 -->
<GroupBox Header="模型" Grid.Row="1" Margin="0,5,0,0">
<GroupBox Header="模型" Grid.Row="2" Margin="0,5,0,0">
<DockPanel>
<StackPanel DockPanel.Dock="Top" Orientation="Horizontal" Margin="0,0,0,5">
<ComboBox x:Name="_modelComboBox"
+47
View File
@@ -29,6 +29,7 @@ namespace AnotherReplayReader
{
InitializeComponent();
_settings = AiSettings.Load();
RefreshPromptFields();
RefreshProviderList();
if (_settings.Providers.Count > 0)
{
@@ -46,6 +47,52 @@ namespace AnotherReplayReader
return new AiRequestContext(_currentProvider, _currentModel);
}
public AiPromptSettings GetPromptSettings()
{
return _settings.Prompt;
}
private void RefreshPromptFields()
{
_useCustomPromptCheck.IsChecked = _settings.Prompt.UseCustomSystemPrompt;
_customPromptBox.Text = _settings.Prompt.CustomSystemPrompt;
_additionalRulesBox.Text = _settings.Prompt.AdditionalRules;
}
private void OnApplyPromptClick(object sender, RoutedEventArgs e)
{
_settings.Prompt.UseCustomSystemPrompt = _useCustomPromptCheck.IsChecked == true;
_settings.Prompt.CustomSystemPrompt = _customPromptBox.Text;
_settings.Prompt.AdditionalRules = _additionalRulesBox.Text;
_settings.Save();
MessageBox.Show("提示词配置已保存", "信息", MessageBoxButton.OK, MessageBoxImage.Information);
}
private void OnResetPromptClick(object sender, RoutedEventArgs e)
{
var result = MessageBox.Show(
"确定要恢复内置 System Prompt 吗?自定义 System Prompt 会被清空,补充规则会保留。",
"恢复默认提示词",
MessageBoxButton.YesNo,
MessageBoxImage.Question);
if (result != MessageBoxResult.Yes)
{
return;
}
_settings.Prompt.UseCustomSystemPrompt = false;
_settings.Prompt.CustomSystemPrompt = string.Empty;
_settings.Save();
RefreshPromptFields();
}
private void OnClearAdditionalRulesClick(object sender, RoutedEventArgs e)
{
_settings.Prompt.AdditionalRules = string.Empty;
_settings.Save();
RefreshPromptFields();
}
// ---------- Provider 列表管理 ----------
private void RefreshProviderList()
{
+21
View File
@@ -0,0 +1,21 @@
# Context
## Glossary
### Operation Fact
A fact directly extracted from replay command data, such as command time, player, command name, UnitId, asset id, special power id, production queue, or target position.
### UnitId Guess
A hypothesis about what game object a UnitId represents. A UnitId Guess is not a fact unless it is directly supported by replay data or game rules.
### Evidence Level
The confidence assigned to a UnitId Guess or tactical conclusion. Valid levels are: confirmed, highly likely, possible, uncertain, and ruled out.
### Validation Rule
A deterministic rule that checks LLM claims against Operation Facts and known game rules.
### Validation Issue
A machine-detected problem in an LLM claim, such as a direct contradiction, weak evidence, missing alternative, or impossible timeline.
### Revision Pass
A hidden LLM request that receives the prior draft, validation issues, and relevant Operation Facts, then produces a corrected analysis without exposing apology or correction chatter to the user.
+1
View File
@@ -449,6 +449,7 @@ namespace AnotherReplayReader
return;
}
_aiPanel.GetRequestContext = _aiSettings.GetCurrentContext!;
_aiPanel.GetPromptSettings = _aiSettings.GetPromptSettings;
await _aiPanel.StartAnalysisAsync(replay, _model.Players, cached, cachedPrefixSums, _cancellation.Token);
}
+381
View File
@@ -0,0 +1,381 @@
using System;
using System.Collections.Generic;
using System.Collections.Immutable;
using System.Linq;
using System.Text;
using System.Text.Json;
using System.Text.RegularExpressions;
namespace AnotherReplayReader.Utils
{
internal enum AIValidationSeverity
{
Info,
WeakEvidence,
Warning,
Contradiction,
Fatal
}
internal enum AIValidationIssueKind
{
InvalidMachineReadableClaims,
MissingMachineReadableClaims,
MissingAlternative,
WeakEvidence,
MissingEvidence,
InvalidEvidenceLevel,
UnitCapabilityContradiction,
UnitTimelineContradiction,
UnsupportedGameKnowledge
}
internal enum AIEvidenceLevel
{
Confirmed,
HighlyLikely,
Possible,
Uncertain,
RuledOut
}
internal sealed record AIUnitClaim(
string UnitId,
string Player,
string Claim,
AIEvidenceLevel EvidenceLevel,
ImmutableArray<string> Evidence,
ImmutableArray<string> Alternatives,
ImmutableArray<string> NeedsConfirmation);
internal sealed record AIEventClaim(
string Claim,
AIEvidenceLevel EvidenceLevel,
ImmutableArray<string> Evidence);
internal sealed record AITimelineClaim(
string Claim,
AIEvidenceLevel EvidenceLevel,
ImmutableArray<string> Evidence);
internal sealed record AIMachineReadableClaims(
ImmutableArray<AIUnitClaim> UnitClaims,
ImmutableArray<AIEventClaim> EventClaims,
ImmutableArray<AITimelineClaim> TimelineClaims)
{
public static AIMachineReadableClaims Empty { get; } = new(
ImmutableArray<AIUnitClaim>.Empty,
ImmutableArray<AIEventClaim>.Empty,
ImmutableArray<AITimelineClaim>.Empty);
}
internal sealed record AIValidationIssue(
AIValidationSeverity Severity,
AIValidationIssueKind Kind,
string Message,
string? UnitId = null,
TimeSpan? Time = null);
internal sealed record AIValidationResult(
AIMachineReadableClaims Claims,
ImmutableArray<AIValidationIssue> Issues)
{
public static AIValidationResult Empty { get; } =
new(AIMachineReadableClaims.Empty, ImmutableArray<AIValidationIssue>.Empty);
public bool RequiresRevision =>
Issues.Any(i => i.Severity is AIValidationSeverity.Contradiction or AIValidationSeverity.Fatal);
public bool HasIssues => !Issues.IsEmpty;
}
internal static class AIAnalysisValidation
{
private static readonly Regex _jsonFenceRegex = new(
@"```(?:json)?\s*(\{[\s\S]*?\})\s*```",
RegexOptions.IgnoreCase | RegexOptions.Compiled);
public static AIValidationResult ValidateMachineReadableClaims(string response)
{
var issues = ImmutableArray.CreateBuilder<AIValidationIssue>();
var json = ExtractJsonObject(response);
if (json is null || string.IsNullOrWhiteSpace(json))
{
issues.Add(new AIValidationIssue(
AIValidationSeverity.Warning,
AIValidationIssueKind.MissingMachineReadableClaims,
"AI 未输出机器可读声明,无法进行自动验证。"));
return new AIValidationResult(AIMachineReadableClaims.Empty, issues.ToImmutable());
}
var claims = ParseClaims(json, issues);
if (claims is null)
{
return new AIValidationResult(AIMachineReadableClaims.Empty, issues.ToImmutable());
}
ValidateClaimSelfConsistency(claims, issues);
return new AIValidationResult(claims, issues.ToImmutable());
}
public static string FormatIssues(ImmutableArray<AIValidationIssue> issues)
{
if (issues.IsEmpty)
{
return "未发现机器可读声明问题。";
}
var sb = new StringBuilder();
foreach (var issue in issues)
{
var unitText = string.IsNullOrWhiteSpace(issue.UnitId)
? string.Empty
: $" UnitId={issue.UnitId}";
sb.AppendLine($"[{issue.Severity}/{issue.Kind}]{unitText} {issue.Message}");
}
return sb.ToString().TrimEnd();
}
private static string? ExtractJsonObject(string response)
{
var markerIndex = response.LastIndexOf("[机器可读声明]", StringComparison.OrdinalIgnoreCase);
var searchText = markerIndex >= 0 ? response.Substring(markerIndex) : response;
var matches = _jsonFenceRegex.Matches(searchText);
if (matches.Count > 0)
{
return matches[matches.Count - 1].Groups[1].Value;
}
var start = searchText.LastIndexOf('{');
var end = searchText.LastIndexOf('}');
if (start >= 0 && end > start)
{
return searchText.Substring(start, end - start + 1);
}
return null;
}
private static AIMachineReadableClaims? ParseClaims(
string json,
ImmutableArray<AIValidationIssue>.Builder issues)
{
try
{
using var doc = JsonDocument.Parse(json, new JsonDocumentOptions
{
AllowTrailingCommas = true,
CommentHandling = JsonCommentHandling.Skip
});
var root = doc.RootElement;
if (root.ValueKind != JsonValueKind.Object)
{
issues.Add(new AIValidationIssue(
AIValidationSeverity.Warning,
AIValidationIssueKind.InvalidMachineReadableClaims,
"机器可读声明不是 JSON object。"));
return null;
}
return new AIMachineReadableClaims(
ReadUnitClaims(root),
ReadSimpleClaims(root, "eventClaims")
.Select(c => new AIEventClaim(c.Claim, c.EvidenceLevel, c.Evidence))
.ToImmutableArray(),
ReadSimpleClaims(root, "timelineClaims")
.Select(c => new AITimelineClaim(c.Claim, c.EvidenceLevel, c.Evidence))
.ToImmutableArray());
}
catch (JsonException ex)
{
issues.Add(new AIValidationIssue(
AIValidationSeverity.Warning,
AIValidationIssueKind.InvalidMachineReadableClaims,
$"机器可读声明 JSON 解析失败:{ex.Message}"));
return null;
}
}
private static ImmutableArray<AIUnitClaim> ReadUnitClaims(JsonElement root)
{
if (!TryGetArray(root, "unitClaims", out var unitClaims))
{
return ImmutableArray<AIUnitClaim>.Empty;
}
var result = ImmutableArray.CreateBuilder<AIUnitClaim>();
foreach (var item in unitClaims.EnumerateArray())
{
if (item.ValueKind != JsonValueKind.Object)
{
continue;
}
result.Add(new AIUnitClaim(
ReadFlexibleString(item, "unitId"),
ReadFlexibleString(item, "player"),
ReadFlexibleString(item, "claim"),
ReadEvidenceLevel(item),
ReadStringArray(item, "evidence"),
ReadStringArray(item, "alternatives"),
ReadStringArray(item, "needsConfirmation")));
}
return result.ToImmutable();
}
private sealed record SimpleClaim(
string Claim,
AIEvidenceLevel EvidenceLevel,
ImmutableArray<string> Evidence);
private static ImmutableArray<SimpleClaim> ReadSimpleClaims(JsonElement root, string propertyName)
{
if (!TryGetArray(root, propertyName, out var claims))
{
return ImmutableArray<SimpleClaim>.Empty;
}
var result = ImmutableArray.CreateBuilder<SimpleClaim>();
foreach (var item in claims.EnumerateArray())
{
if (item.ValueKind != JsonValueKind.Object)
{
continue;
}
result.Add(new SimpleClaim(
ReadFlexibleString(item, "claim"),
ReadEvidenceLevel(item),
ReadStringArray(item, "evidence")));
}
return result.ToImmutable();
}
private static void ValidateClaimSelfConsistency(
AIMachineReadableClaims claims,
ImmutableArray<AIValidationIssue>.Builder issues)
{
foreach (var claim in claims.UnitClaims)
{
if (string.IsNullOrWhiteSpace(claim.UnitId))
{
issues.Add(new AIValidationIssue(
AIValidationSeverity.Warning,
AIValidationIssueKind.InvalidMachineReadableClaims,
"Unit claim 缺少 unitId。"));
}
if (string.IsNullOrWhiteSpace(claim.Claim))
{
issues.Add(new AIValidationIssue(
AIValidationSeverity.Warning,
AIValidationIssueKind.InvalidMachineReadableClaims,
"Unit claim 缺少 claim。",
claim.UnitId));
}
if (claim.EvidenceLevel is AIEvidenceLevel.Confirmed or AIEvidenceLevel.HighlyLikely
&& claim.Evidence.IsEmpty)
{
issues.Add(new AIValidationIssue(
AIValidationSeverity.Warning,
AIValidationIssueKind.MissingEvidence,
"高置信 UnitId 推测缺少 evidence。",
claim.UnitId));
}
if (claim.EvidenceLevel is AIEvidenceLevel.Possible or AIEvidenceLevel.Uncertain
&& claim.Alternatives.IsEmpty
&& claim.NeedsConfirmation.IsEmpty)
{
issues.Add(new AIValidationIssue(
AIValidationSeverity.WeakEvidence,
AIValidationIssueKind.MissingAlternative,
"低置信 UnitId 推测应提供 alternatives 或 needsConfirmation。",
claim.UnitId));
}
}
}
private static bool TryGetArray(JsonElement root, string propertyName, out JsonElement array)
{
if (root.TryGetProperty(propertyName, out array)
&& array.ValueKind == JsonValueKind.Array)
{
return true;
}
array = default;
return false;
}
private static AIEvidenceLevel ReadEvidenceLevel(JsonElement item)
{
var value = ReadFlexibleString(item, "evidenceLevel");
return NormalizeEvidenceLevel(value);
}
private static AIEvidenceLevel NormalizeEvidenceLevel(string value)
{
value = value.Trim().Replace("_", "").Replace("-", "").Replace(" ", "");
return value.ToLowerInvariant() switch
{
"confirmed" or "确定" => AIEvidenceLevel.Confirmed,
"highlylikely" or "high" or "高度可能" => AIEvidenceLevel.HighlyLikely,
"possible" or "可能" => AIEvidenceLevel.Possible,
"ruledout" or "excluded" or "已排除" => AIEvidenceLevel.RuledOut,
_ => AIEvidenceLevel.Uncertain,
};
}
private static string ReadFlexibleString(JsonElement item, string propertyName)
{
if (!item.TryGetProperty(propertyName, out var value))
{
return string.Empty;
}
return value.ValueKind switch
{
JsonValueKind.String => value.GetString() ?? string.Empty,
JsonValueKind.Number => value.GetRawText(),
JsonValueKind.True => "true",
JsonValueKind.False => "false",
_ => string.Empty,
};
}
private static ImmutableArray<string> ReadStringArray(JsonElement item, string propertyName)
{
if (!item.TryGetProperty(propertyName, out var value))
{
return ImmutableArray<string>.Empty;
}
if (value.ValueKind == JsonValueKind.String)
{
var text = value.GetString();
return string.IsNullOrWhiteSpace(text)
? ImmutableArray<string>.Empty
: ImmutableArray.Create(text!);
}
if (value.ValueKind != JsonValueKind.Array)
{
return ImmutableArray<string>.Empty;
}
var result = new List<string>();
foreach (var element in value.EnumerateArray())
{
var text = element.ValueKind == JsonValueKind.String
? element.GetString()
: element.GetRawText();
if (!string.IsNullOrWhiteSpace(text))
{
result.Add(text!);
}
}
return result.ToImmutableArray();
}
}
}
+67 -3
View File
@@ -17,13 +17,51 @@ namespace AnotherReplayReader.Utils
{
internal sealed class AIAnalyze
{
public static string GetSystemPrompt(Replay replay, ImmutableSortedDictionary<int, Player> players)
public static string GetSystemPrompt(
Replay replay,
ImmutableSortedDictionary<int, Player> players,
AiPromptSettings? promptSettings = null)
{
var defaultPrompt = BuildDefaultSystemPrompt(replay, players);
return ComposeSystemPrompt(defaultPrompt, promptSettings);
}
public static string ComposeSystemPrompt(string defaultPrompt, AiPromptSettings? promptSettings)
{
var customSystemPrompt = promptSettings?.CustomSystemPrompt ?? string.Empty;
if (promptSettings?.UseCustomSystemPrompt == true
&& !string.IsNullOrWhiteSpace(customSystemPrompt))
{
defaultPrompt = customSystemPrompt;
}
var additionalRules = promptSettings?.AdditionalRules ?? string.Empty;
if (!string.IsNullOrWhiteSpace(additionalRules))
{
defaultPrompt = defaultPrompt.TrimEnd()
+ "\n\n# 用户自定义补充规则\n"
+ additionalRules.Trim();
}
return defaultPrompt.Replace("\r", "");
}
public static string BuildDefaultSystemPrompt(Replay replay, ImmutableSortedDictionary<int, Player> players)
{
var generalDescriptions = @"
你是一位 RTS 游戏数据分析师,你擅长从大量数据中发现有趣的规律和细节。
用户则是一位玩家,用户会向你提供玩家操作记录,你要对其进行分析。
# 核心原则
- 你只能根据用户提供的操作记录、玩家信息、下方游戏规则和明确给出的背景知识进行分析。
- 不要使用现实世界常识或其他 RTS 游戏常识覆盖这里的游戏设定。例如:步兵、直升机、建筑水陆摆放、运输能力、两栖能力都必须以这里的规则和单位描述为准。
- 不确定时必须保留多个候选,不要为了让解说流畅而过早下定论。
- 对 UnitId、单位类型、战术意图的判断必须区分证据等级:确定、高度可能、可能、待确认、已排除。
- 每个关键推理都应当包含支持证据;如果存在会推翻该推理的反证,也要主动指出。
- 如果某个技能或行为可以对应多个单位,先列出候选,并说明还需要哪些后续迹象才能确认。
- 对已经被操作记录直接否定的判断必须修正或放弃,不要坚持原结论。
# 输入格式
## 用户初始输入
- 玩家信息
@@ -83,11 +121,36 @@ namespace AnotherReplayReader.Utils
- 也可以重点关注PlayerTech、英雄、工程师
- 按照**推理指南**进行详细的思考与推理,列举你的推理与发现
- 输出:该阶段的各个主要事件,以及你的推理和发现
- 假如推测 UnitId 对应的单位,请在正文中自然描述,并在末尾输出机器可读声明,方便程序验证
## 3. 最终总结阶段
触发条件:用户输入包含:""请对以上内容进行总结""
- 输出:所有分析的总结,以及这次对局的完整介绍
# 机器可读声明
仅限于:分段分析阶段(第2阶段)
如果你对 UnitId、关键事件或时间线做出了可验证推测,请在回答末尾附加下面格式。
必须先输出一行`[机器可读声明]`,然后输出一个 JSON 代码块:
[机器可读声明]
```json
{
""unitClaims"": [
{
""unitId"": 123,
""player"": ""PlayerA"",
""claim"": ""AlliedMCV"",
""evidenceLevel"": ""possible"",
""evidence"": [""8:30 使用 SpecialPower_UnpackReplaceSelf""],
""alternatives"": [""AlliedMiner 展开后的指挥中心""],
""needsConfirmation"": [""是否曾使用 SpecialPower_PackReplaceSelf"", ""后续是否作为建造者出现""]
}
],
""eventClaims"": [],
""timelineClaims"": []
}
```
如果没有可验证推测,可以输出空数组。不要在 JSON 里写注释。
# 推理指南
推理需要分成多个阶段
1. 观察
@@ -953,6 +1016,7 @@ PlayerA: 开始出兵
请按照按照[观察]、[分析]、[推理]、[进一步思考(可选)]的步骤,对各个事件进行分析和推理。
假如当前阶段存在一些较为重要的单位、而且能够推测出它们可能是什么单位,则可以列出单位的UnitId以及你对单位的推测
如果你列出了 UnitId 推测、关键事件推测或时间线推测,请在回答末尾附加`[机器可读声明]` JSON 代码块;如果没有相关推测,则输出空数组。
";
return instruction.Trim().Replace("\r", "");
}
@@ -1273,8 +1337,8 @@ PlayerA: 开始出兵
result.ReasoningTokens = GetIntegerProperty(usage, "reasoning_tokens") ?? result.ReasoningTokens;
}
// 提取 usage(如果存在)
if (doc.RootElement.TryGetProperty("error", out var error) && usage.ValueKind == JsonValueKind.Object)
// 提取 error(如果存在)
if (doc.RootElement.TryGetProperty("error", out var error) && error.ValueKind == JsonValueKind.Object)
{
if (error.TryGetProperty("message", out var message))
{
+13 -1
View File
@@ -18,11 +18,21 @@ namespace AnotherReplayReader
public string ApiKey { get; set; } = string.Empty;
public List<AiModel> Models { get; set; } = [];
public double DefaultTemperature { get; set; } = 0.75;
public double DefaultTemperature { get; set; } = 0.35;
public double DefaultTopP { get; set; } = 0.95;
public int DefaultMaxTokens { get; set; } = 16384;
}
/// <summary>
/// AI 分析提示词配置。
/// </summary>
public class AiPromptSettings
{
public bool UseCustomSystemPrompt { get; set; }
public string CustomSystemPrompt { get; set; } = string.Empty;
public string AdditionalRules { get; set; } = string.Empty;
}
/// <summary>
/// 模型条目
/// </summary>
@@ -270,6 +280,7 @@ namespace AnotherReplayReader
public class AiSettings
{
public List<AiProvider> Providers { get; set; } = [];
public AiPromptSettings Prompt { get; set; } = new();
// 以下两个不持久化,由 UI 层维护当前选中项
[System.Text.Json.Serialization.JsonIgnore]
@@ -289,6 +300,7 @@ namespace AnotherReplayReader
var settings = JsonSerializer.Deserialize<AiSettings>(json);
if (settings is { } value && value.Providers.Count > 0)
{
value.Prompt ??= new AiPromptSettings();
return settings;
}
}
+233
View File
@@ -0,0 +1,233 @@
# AI Analysis WIP
## User Need
The application is adding AI analysis for Red Alert 3 replay operation logs. The current flow sends player information and a compacted operation log to a chat-completion-compatible LLM, then displays the analysis in `AIChatPanel`.
The main goals are:
- Improve the current system prompt.
- Make the system prompt configurable by the user.
- Reduce AI analysis errors, especially errors caused by general-world assumptions or overconfident UnitId guesses.
- Add a validation path for LLM output so wrong claims can be detected and corrected without wasting all prior reasoning.
## Current Code Areas
- `Utils/AIAnalyze.cs`: builds the system prompt, user prompts, segment prompts, final summary prompts, and performs OpenAI-compatible chat completion calls.
- `AIChatPanel.xaml.cs`: runs the analysis workflow, displays streaming chunks, retries failed segments, and now logs machine-readable claim validation results.
- `Utils/AiSettings.cs`: stores AI provider/model settings and now prompt settings.
- `AIProviderSettingsControl.xaml(.cs)`: edits provider/model settings and now prompt settings.
- `EventDump.xaml.cs`: generates the replay operation text and starts AI analysis.
- `Utils/AIAnalysisValidation.cs`: new validation model and parser for machine-readable AI claims.
- `CONTEXT.md`: glossary for the AI analysis domain.
## Discussion Notes
### Prompt Problems
The current prompt has a lot of useful game knowledge, but the LLM can still:
- Use common sense that is wrong for the game or mod.
- Assume infantry, helicopters, transports, amphibious movement, and water placement work like they do in other RTS games.
- Treat one observed skill as conclusive evidence when multiple units share that skill.
- Overstate UnitId guesses.
Examples discussed:
- Only units explicitly marked amphibious can move on both land and water.
- Only units explicitly marked as passenger transports can transport infantry.
- Building water placement depends on game rules, not common assumptions.
- `SpecialPower_UnpackReplaceSelf` does not uniquely identify an Allied MCV because Allied miners can also unpack into a command hub.
- A UnitId claimed as an aircraft should be challenged if the same UnitId is observed using an unpack/deploy skill.
- A UnitId claimed as a bomber should be challenged if it is operated before the player starts producing their first bomber.
### Prompt Decisions
The default prompt should explicitly require:
- Evidence-first analysis.
- No use of external common sense over replay facts and supplied game rules.
- UnitId guesses with evidence levels.
- Multiple candidates when a behavior has several possible sources.
- Support evidence and possible counter-evidence for important claims.
- Correction or abandonment of claims contradicted by replay facts.
Evidence levels currently used:
- confirmed
- highly likely
- possible
- uncertain
- ruled out
The default provider temperature was lowered from `0.75` to `0.35` because this task is closer to audit/reconstruction than creative writing.
### Prompt Configuration Decisions
The prompt is now configurable through AI settings.
The design has two prompt layers:
- A built-in dynamic system prompt, still assembled from replay/mod/faction/map context.
- User prompt settings:
- optional full custom system prompt
- additional rules appended to the final system prompt
This keeps the normal path safe while allowing advanced users to override the whole prompt.
### Validation Philosophy
LLM natural-language analysis should not be treated as directly valid. The plan is to validate structured claims emitted by the LLM.
Important decision:
- Do not immediately throw away a whole analysis when a problem is found.
- Do not show the user two competing analyses or apology text such as "sorry, my previous answer was wrong."
- Prefer a hidden revision pass: send the draft, validation issues, and relevant replay facts back to the AI, asking it to output a clean corrected version without mentioning the revision.
- Limit retries/revisions. If the model still cannot resolve a claim, downgrade confidence or mark it uncertain instead of looping forever.
Severity model:
- `Info`: useful diagnostic only.
- `WeakEvidence`: claim may be plausible but lacks enough support.
- `Warning`: malformed or questionable claim that should be logged or possibly revised.
- `Contradiction`: claim conflicts with replay facts or game rules and should trigger revision.
- `Fatal`: output cannot be used for the current phase, such as empty or unparseable required output.
### JSON Format Decision
We discussed whether to require JSON or use a simpler line-based format.
Decision:
- Use JSON for machine-readable claims.
- Keep the schema small.
- Make the parser tolerant.
Reasoning:
- JSON can naturally represent evidence arrays, alternatives, and needed confirmations.
- A custom line format would be easier for a trivial parser but would become fragile once nested data is needed.
- The app can tolerate partial or missing JSON by logging validation issues instead of failing the whole analysis.
Current expected shape:
```json
{
"unitClaims": [
{
"unitId": 123,
"player": "PlayerA",
"claim": "AlliedMCV",
"evidenceLevel": "possible",
"evidence": ["8:30 使用 SpecialPower_UnpackReplaceSelf"],
"alternatives": ["AlliedMiner 展开后的指挥中心"],
"needsConfirmation": ["是否曾使用 SpecialPower_PackReplaceSelf", "后续是否作为建造者出现"]
}
],
"eventClaims": [],
"timelineClaims": []
}
```
The prompt asks the AI to output:
```text
[机器可读声明]
```json
{ ... }
```
```
The parser first looks for the last fenced JSON block near `[机器可读声明]`, then falls back to the last `{...}` block.
## Validation We Can Do
### Implemented Now
Format validation:
- Missing machine-readable claims.
- JSON parse failure.
- Root value is not an object.
Self-consistency validation:
- Unit claim missing `unitId`.
- Unit claim missing `claim`.
- High-confidence UnitId guess without evidence.
- Low-confidence UnitId guess without alternatives or needed confirmation.
### Near-Term Validations
These need replay facts extracted from `CommandChunk` or an intermediate fact index:
- UnitId special power contradictions.
- UnitId production timeline contradictions.
- UnitId used as builder vs claimed as non-builder unit.
- Claims that use game knowledge not present in rules, such as transport/amphibious/building-placement abilities.
- Missing alternatives for ambiguous skills such as Allied unpack.
### Suggested Fact Index
Useful derived facts:
- `UnitId -> first observed time`
- `UnitId -> observed special powers`
- `UnitId -> observed as builder`
- `UnitId -> observed as production structure`
- `Player -> first production time by asset id`
- `Player -> selected UnitIds over time`
- `Player -> tech/protocol choices`
- `Player -> building placements by asset and position`
## Current Progress
Implemented:
- Added `CONTEXT.md` glossary.
- Added prompt settings:
- `AiPromptSettings`
- `UseCustomSystemPrompt`
- `CustomSystemPrompt`
- `AdditionalRules`
- Added prompt editing UI to `AIProviderSettingsControl`.
- Connected prompt settings from `EventDump` to `AIChatPanel` to `AIAnalyze`.
- Split default prompt construction from prompt composition.
- Strengthened default prompt with evidence-first and uncertainty rules.
- Added machine-readable JSON claim instructions to system and segment prompts.
- Added `Utils/AIAnalysisValidation.cs` with:
- evidence level enum
- machine-readable claim records
- validation issue records
- JSON extraction and parsing
- initial self-consistency checks
- Added per-segment validation logging in `AIChatPanel`.
Build status:
- `dotnet build AnotherReplayReader.csproj --no-restore` succeeds.
- Remaining warnings are existing nullable warnings in `AIAnalyze.cs` stream response handling and a `System.Text.Encoding.CodePages` support warning for `net461`.
## Open Questions
- Should AI natural-language output continue streaming live, or should content be buffered until validation and possible revision are complete?
- Should reasoning chunks remain visible during hidden revision, or should only final content be shown?
- How strict should missing machine-readable claims be?
- Current behavior: warning log only.
- Possible future behavior: one hidden repair request asking the model to append valid claims.
- Should validation issues be visible by default, or only in an advanced/debug foldout?
- How much game-unit knowledge should live in code versus prompt text?
- Should the first verifier use hardcoded RA3/Corona knowledge, or should it load a small unit capability table from data files?
## Suggested Next Steps
1. Build a replay fact index from `CommandChunk`.
2. Add first deterministic validation rules:
- ambiguous Allied unpack
- pack/unpack consistency
- UnitId used as builder
- first production time vs first operation time
3. Decide whether to buffer per-segment content before display.
4. Add one hidden revision pass for `Contradiction` issues.
5. Add validation summary UI, such as "验证器发现并修正 N 个问题".