Update token allocation for DeepSeek R1 128K context window - Set NOVITA_MODEL_CONTEXT_WINDOW to 128000 (128K tokens) - Increase USER_INPUT_MAX_TOKENS to 32000 - Increase CONTEXT_PREPARATION_BUDGET to 115000 - Increase CONTEXT_PRUNING_THRESHOLD to 115000 - Update validators to support larger token limits

ENV_EXAMPLE_CONTENT.txt

@@ -31,23 +31,24 @@ DEEPSEEK_R1_FORCE_REASONING=True
 # Token Allocation Configuration
 # =============================================================================
 # Maximum tokens dedicated for user input (prioritized over context)
-# Recommended: 8000 tokens for large queries
-USER_INPUT_MAX_TOKENS=8000
+# Recommended: 32000 tokens for DeepSeek R1 (128K context window)
+USER_INPUT_MAX_TOKENS=32000
 
 # Maximum tokens for context preparation (includes user input + context)
-# Recommended: 28000 tokens for 32K context window models
-CONTEXT_PREPARATION_BUDGET=28000
+# Recommended: 115000 tokens for DeepSeek R1 (leaves ~13K for output)
+CONTEXT_PREPARATION_BUDGET=115000
 
 # Context pruning threshold (should match context_preparation_budget)
-CONTEXT_PRUNING_THRESHOLD=28000
+CONTEXT_PRUNING_THRESHOLD=115000
 
 # Always prioritize user input over historical context
 PRIORITIZE_USER_INPUT=True
 
 # Model context window (actual limit for your deployed model)
-# Default: 5000 tokens (adjust based on your Novita AI deployment)
+# Default: 128000 tokens for DeepSeek R1 (128K context window)
 # This is the maximum total tokens (input + output) the model can handle
-NOVITA_MODEL_CONTEXT_WINDOW=5000
+# Take full advantage of DeepSeek R1's 128K capability
+NOVITA_MODEL_CONTEXT_WINDOW=128000
 
 # =============================================================================
 # Database Configuration

src/config.py

@@ -209,19 +209,19 @@ class Settings(BaseSettings):
     
     # Token Allocation Configuration
     user_input_max_tokens: int = Field(
-        default=8000,
+        default=32000,
         description="Maximum tokens dedicated for user input (prioritized over context)",
         env="USER_INPUT_MAX_TOKENS"
     )
     
     context_preparation_budget: int = Field(
-        default=28000,
+        default=115000,
         description="Maximum tokens for context preparation (includes user input + context)",
         env="CONTEXT_PREPARATION_BUDGET"
     )
     
     context_pruning_threshold: int = Field(
-        default=28000,
+        default=115000,
         description="Context pruning threshold (should match context_preparation_budget)",
         env="CONTEXT_PRUNING_THRESHOLD"
     )
@@ -234,8 +234,8 @@ class Settings(BaseSettings):
     
     # Model Context Window Configuration
     novita_model_context_window: int = Field(
-        default=5000,
-        description="Maximum context window for Novita AI model (input + output tokens)",
+        default=128000,
+        description="Maximum context window for Novita AI model (input + output tokens). DeepSeek R1 supports 128K tokens.",
         env="NOVITA_MODEL_CONTEXT_WINDOW"
     )
     
@@ -264,20 +264,26 @@ class Settings(BaseSettings):
     @validator("user_input_max_tokens", pre=True)
     def validate_user_input_tokens(cls, v):
         """Validate user input token limit"""
-        val = int(v) if v else 8000
-        return max(1000, min(20000, val))
+        val = int(v) if v else 32000
+        return max(1000, min(50000, val))  # Allow up to 50K for large inputs
     
     @validator("context_preparation_budget", pre=True)
     def validate_context_budget(cls, v):
         """Validate context preparation budget"""
-        val = int(v) if v else 28000
-        return max(4000, min(120000, val))
+        val = int(v) if v else 115000
+        return max(4000, min(125000, val))  # Allow up to 125K for 128K context window
+    
+    @validator("context_pruning_threshold", pre=True)
+    def validate_pruning_threshold(cls, v):
+        """Validate context pruning threshold"""
+        val = int(v) if v else 115000
+        return max(4000, min(125000, val))  # Match context_preparation_budget limits
     
     @validator("novita_model_context_window", pre=True)
     def validate_context_window(cls, v):
         """Validate context window size"""
-        val = int(v) if v else 5000
-        return max(1000, min(200000, val))  # Reasonable bounds
+        val = int(v) if v else 128000
+        return max(1000, min(200000, val))  # Support up to 200K for future models
     
     # ==================== Model Configuration ====================