add reason_prefill to force model to start with thinking block

src/talemate/client/base.py

@@ -96,6 +96,7 @@ class CommonDefaults(pydantic.BaseModel):
     reason_enabled: bool = False
     reason_tokens: int = 0
     reason_response_pattern: str | None = None
+    reason_prefill: str | None = None
 
 
 class Defaults(CommonDefaults, pydantic.BaseModel):
@@ -322,6 +323,10 @@ class ClientBase:
     @property
     def reason_response_pattern(self) -> str:
         return self.client_config.reason_response_pattern or DEFAULT_REASONING_PATTERN
+    
+    @property
+    def reason_prefill(self) -> str:
+        return self.client_config.reason_prefill or ""
 
     @property
     def lock_template(self) -> bool:
@@ -509,6 +514,14 @@ class ClientBase:
             double_coercion = f"{double_coercion}\n\n"
         else:
             double_coercion = None
+            
+        if self.reason_enabled and self.reason_prefill:
+            # if reasoning is enabled and a reason prefill is set, prepend it to the double coercion
+            # its important that it comes first.
+            if not double_coercion:
+                double_coercion = self.reason_prefill
+            else:
+                double_coercion = f"{self.reason_prefill}{double_coercion}"
 
         spec = PromptSpec()
 
@@ -769,6 +782,7 @@ class ClientBase:
             "reason_tokens": self.reason_tokens,
             "min_reason_tokens": self.min_reason_tokens,
             "reason_response_pattern": self.reason_response_pattern,
+            "reason_prefill": self.reason_prefill,
             "requires_reasoning_pattern": self.requires_reasoning_pattern,
             "request_information": self.request_information.model_dump()
             if self.request_information

src/talemate/config/schema.py

@@ -60,6 +60,10 @@ class Client(pydantic.BaseModel):
 
     # regex to strip from the response if the model is reasoning
     reason_response_pattern: Union[str, None] = None
+    
+    # reason prefill - will be prepended to the prompt if the model is reasoning
+    # this is mostly for base models that don't hhave reas
+    reason_prefill: str | None = None
 
     system_prompts: SystemPrompts = SystemPrompts()
 

talemate_frontend/src/components/AIClient.vue

@@ -465,6 +465,7 @@ export default {
           client.reason_tokens = data.data.reason_tokens;
           client.min_reason_tokens = data.data.min_reason_tokens;
           client.reason_response_pattern = data.data.reason_response_pattern;
+          client.reason_prefill = data.data.reason_prefill;
           client.reason_enabled = data.data.reason_enabled;
           client.requires_reasoning_pattern = data.data.requires_reasoning_pattern;
           client.lock_template = data.data.lock_template;
@@ -517,6 +518,7 @@ export default {
             reason_tokens: data.data.reason_tokens,
             min_reason_tokens: data.data.min_reason_tokens,
             reason_response_pattern: data.data.reason_response_pattern,
+            reason_prefill: data.data.reason_prefill,
             reason_enabled: data.data.reason_enabled,
             requires_reasoning_pattern: data.data.requires_reasoning_pattern,
             dedicated_default_template: data.data.dedicated_default_template,

talemate_frontend/src/components/ClientModal.vue

@@ -176,6 +176,14 @@
                       <v-text-field v-model="client.reason_response_pattern" label="Pattern to strip from the response if the model is reasoning" hint="This is a regular expression that will be used to strip out the thinking tokens from the response." placeholder=".*?</think>"></v-text-field>
                     </v-col>
                   </v-row>
+                  <v-row v-if="client.reason_enabled">
+                    <v-col cols="12">
+                      <v-text-field v-model="client.reason_prefill" label="Reason Prefill"></v-text-field>
+                      <v-alert color="muted" variant="text" class="text-caption">
+                        This is mostly for base models that don't have reasoning built in, but were fine-tuned for reasoning. For example add <code class="text-primary">&lt;think&gt;</code> here to force the model to reason. Assuming <code class="text-primary">&lt;think&gt;</code> is the actual start of the thinking process, this may vary depending on the model.
+                      </v-alert>
+                    </v-col>
+                  </v-row>
                 </v-window-item>
                 <!-- SYSTEM PROMPTS -->
                 <v-window-item value="system_prompts">
@@ -406,6 +414,7 @@ export default {
         this.client.reason_tokens = defaults.reason_tokens || 0;
         this.client.min_reason_tokens = defaults.min_reason_tokens || 0;
         this.client.reason_response_pattern = defaults.reason_response_pattern || null;
+        this.client.reason_prefill = defaults.reason_prefill || null;
         this.client.requires_reasoning_pattern = defaults.requires_reasoning_pattern || false;
         this.client.lock_template = defaults.lock_template || false;
         this.client.template_file = defaults.template_file || null;