src/NexusReader.Infrastructure/Services/KnowledgeService.cs

@@ -31,7 +31,7 @@ public class KnowledgeService : IKnowledgeService
     private readonly Tokenizer _tokenizer;
     private readonly ILogger<KnowledgeService> _logger;
     private const string PromptVersion = "1.3";
-    private static readonly ConcurrentDictionary<string, Task<Result<KnowledgePacket>>> _activeRequests = new();
+    private static readonly ConcurrentDictionary<string, Lazy<Task<Result<KnowledgePacket>>>> _activeRequests = new();
 
     public KnowledgeService(
         IChatClient chatClient,
@@ -100,10 +100,38 @@ public class KnowledgeService : IKnowledgeService
 
         // Deduplicate concurrent active requests for the exact same hash
         var requestKey = $"{tenantId}:{hash}:{traceType}";
-        var task = _activeRequests.GetOrAdd(requestKey, _ => 
-            ExecuteAiRequestAndCacheAsync(normalizedText, tenantId, systemPrompt, traceType, ebookId, hash));
+        
+        var lazyTask = _activeRequests.GetOrAdd(requestKey, k => 
+            new Lazy<Task<Result<KnowledgePacket>>>(
+                () => ExecuteAiRequestAndCacheAsync(normalizedText, tenantId, systemPrompt, traceType, ebookId, hash),
+                System.Threading.LazyThreadSafetyMode.ExecutionAndPublication
+            ));
 
-        return await task;
+        try
+        {
+            var result = await lazyTask.Value;
+
+            // If the AI call returned a failure, remove it from the active dictionary 
+            // so subsequent retries have a chance to request the AI again.
+            if (result.IsFailed)
+            {
+                _activeRequests.TryRemove(requestKey, out _);
+            }
+
+            return result;
+        }
+        catch (Exception)
+        {
+            // Evict from active dictionary on hard exceptions to ensure system resiliency
+            _activeRequests.TryRemove(requestKey, out _);
+            throw;
+        }
+        finally
+        {
+            // Once a task successfully finishes and persists to the Persistent Database Cache,
+            // we evict it from RAM (_activeRequests) since future hits will leverage the DB cache.
+            _activeRequests.TryRemove(requestKey, out _);
+        }
     }
 
     private async Task<Result<KnowledgePacket>> ExecuteAiRequestAndCacheAsync(