​

Dual Chat (Arena Mode) Stable

​

Authentication Required

sub | ClaimTypes.NameIdentifier → User UUID (required)
email | ClaimTypes.Email → User email
full_name | name | ClaimTypes.Name → Display name

​

Request Body

if (request == null || string.IsNullOrWhiteSpace(request.Prompt))
    return BadRequest("INVALID_REQUEST");

if (request.SelectionMode == "topper") {
    var pair = await _leaderboardModelSelector.GetTopperAndRandomModelAsync();
} else {
    var pair = await _modelSelector.GetTwoRandomModelsAsync();
}

if (manual && (string.IsNullOrWhiteSpace(request.Model1) || 
                string.IsNullOrWhiteSpace(request.Model2))) {
    return BadRequest("Both model1 and model2 are required");
}

var sessionId = Guid.NewGuid();

if (!string.IsNullOrEmpty(request.ThreadId)) {
    if (Guid.TryParse(request.ThreadId, out Guid threadIdGuid)) {
        await _threadMessagesService.LogDualAsync(...);
    }
}

​

Response

if (msg1Len > msg2Len) return "agent1";
else if (msg2Len > msg1Len) return "agent2";
else return "tie";

curl -X POST 'http://localhost:5079/api/arena/dualchat' \
  -H 'Authorization: Bearer YOUR_JWT_TOKEN' \
  -H 'Content-Type: application/json' \
  -d '{
    "prompt": "Write a haiku about programming",
    "selectionMode": "random",
    "temperature": 0.9
  }'

{
  "success": true,
  "agent1": {
    "object": "ai.response",
    "message": "Code flows like water\nThrough circuits of logic pure\nBugs hide in shadows",
    "model": {
      "name": "llama-3.3-70b-versatile",
      "displayName": "Llama 3.3 70B",
      "provider": "groq"
    },
    "usage": {
      "promptTokens": 12,
      "completionTokens": 25,
      "totalTokens": 37
    },
    "responseTimeMs": 823,
    "timestamp": "2024-01-15T10:30:00.000Z"
  },
  "agent2": {
    "object": "ai.response",
    "message": "Silent keystrokes fall\nAlgorithms come alive\nCreation awaits",
    "model": {
      "name": "mixtral-8x7b-32768",
      "displayName": "Mixtral 8x7B",
      "provider": "groq"
    },
    "usage": {
      "promptTokens": 12,
      "completionTokens": 22,
      "totalTokens": 34
    },
    "responseTimeMs": 1102,
    "timestamp": "2024-01-15T10:30:00.000Z"
  },
  "comparisonId": "c7d3a4b2-9e1f-4c5d-8b3a-7f6e9d2c1a0b",
  "arena": {
    "comparison": {
      "winnerByLength": "agent1",
      "winnerByTokens": "agent1",
      "verdict": "Agent 1 produced the longer, more token-heavy answer.",
      "userWinner": null,
      "agent1MessageLength": 68,
      "agent2MessageLength": 55,
      "agent1Tokens": 37,
      "agent2Tokens": 34
    },
    "models": {
      "agent1": "llama-3.3-70b-versatile",
      "agent2": "mixtral-8x7b-32768"
    }
  },
  "timestamp": "2024-01-15T10:30:00.000Z",
  "totalResponseTimeMs": 1102
}

​

Side Effects

1. message_logs table (Lines 333-334):
   await _messageLogger.LogMessageAsync(sessionId, finalModel1, "agent1", request, response1);
   await _messageLogger.LogMessageAsync(sessionId, finalModel2, "agent2", request, response2);
   
   • Two separate log entries (one per model)
2. users table UPSERT (Lines 336-350):
   await _userSyncService.EnsureUserExistsAsync(userId, email, name);
   
   • Executes on every authenticated request
   • Idempotent operation
3. comparisons table (Line 352):
   await _comparisonLogger.LogComparisonAsync(comparisonId, request, response1, response2, userId);
   
   • Stores comparison data for voting/leaderboard
   • Links to both models
4. thread_messages table (conditional, Lines 354-360):
   if (!string.IsNullOrEmpty(request.ThreadId)) {
       await _threadMessagesService.LogDualAsync(threadIdGuid, request.Prompt, 
           finalModel1, finalModel2, response1, response2, comparisonId);
   }
   
   • Only if threadId provided and valid GUID
   • Links message to comparison via comparisonId

​

Behavior

var task1 = ExecuteWithFallbackAsync(model1, ...);
var task2 = ExecuteWithFallbackAsync(model2, ...);

await Task.WhenAll(task1, task2);

• Each model has independent 45s timeout
• Each model has independent fallback chain
• One model failure doesn’t block the other

var responseTime = (long)(DateTime.UtcNow - startTime).TotalMilliseconds;

• Measures total elapsed time
• Due to parallel execution: max(model1_time, model2_time) + overhead

var pair = await _leaderboardModelSelector.GetTopperAndRandomModelAsync();

• Queries model_votes table for highest win rate
• Pairs top model with random model
• Ensures diverse comparison

var msg1Len = (response1.Message ?? string.Empty).Length;
var msg2Len = (response2.Message ?? string.Empty).Length;
var tokens1 = response1.Usage?.TotalTokens ?? 0;
var tokens2 = response2.Usage?.TotalTokens ?? 0;

// Determine winners by length and tokens
// Generate verdict based on combination

​

Error Conditions

try {
    // Dual chat logic
} catch (Exception ex) {
    _logger.LogError(ex, "DualChat inner error");
    return 500 API_ERROR;
}
} catch (Exception outerEx) {
    _logger.LogError(outerEx, "DualChat outer error");
    return 500 API_ERROR;
}

​

Edge Cases

1. Same model selected twice: Not prevented by code, allowed in random selection
2. Topper mode with insufficient vote data: Behavior not enforced by server contract (assumed fallback to random selection)
3. Invalid threadId GUID: Silently skipped, no error (Line 356 guard)
4. Model fallback changes model names: finalModel1 and finalModel2 may differ from requested models
5. Null usage stats: Handled with null-coalescing (Lines 366-367)

​

Comparison ID Usage

var comparisonId = Guid.NewGuid();

1. Logging comparison to comparisons table (Line 352)
2. Linking to thread message in thread_messages table (Line 358)
3. Returned in response for voting (Line 395)
4. Voting endpoint requires this ID: POST /api/arena/model-vote

​

Rate Limits

• Groq free tier: 30 req/min, 14,400 tokens/min
• Dual chat consumes 2× tokens (both models)
• Effective limit: ~15 dual-chat requests/min on free tier