FEATURE: Select stop sequences from triage script

Author: romanrizzi

config/locales/client.en.yml

@@ -89,6 +89,9 @@ en:
             max_post_tokens:
               label: "Max Post Tokens"
               description: "The maximum number of tokens to scan using LLM triage"
+            stop_secuences:
+              label: "Stop Secuences"
+              description: "Instruct the model to halt token generation when arriving at one of these values"
             search_for_text:
               label: "Search for text"
               description: "If the following text appears in the LLM reply, apply these actions"

discourse_automation/llm_triage.rb

@@ -12,6 +12,7 @@
     field :system_prompt, component: :message, required: false
     field :search_for_text, component: :text, required: true
     field :max_post_tokens, component: :text
+    field :stop_sequences, component: :text_list, required: false
     field :model,
           component: :choices,
           required: true,
@@ -55,6 +56,8 @@
 
       max_post_tokens = nil if max_post_tokens <= 0
 
+      stop_sequences = fields.dig("stop_sequences", "value")
+
       if post.topic.private_message?
         include_personal_messages = fields.dig("include_personal_messages", "value")
         next if !include_personal_messages
@@ -88,6 +91,7 @@
           flag_post: flag_post,
           flag_type: flag_type.to_s.to_sym,
           max_post_tokens: max_post_tokens,
+          stop_sequences: stop_sequences,
           automation: self.automation,
         )
       rescue => e

lib/automation/llm_triage.rb

@@ -16,7 +16,8 @@ def self.handle(
         flag_post: nil,
         flag_type: nil,
         automation: nil,
-        max_post_tokens: nil
+        max_post_tokens: nil,
+        stop_sequences: nil
       )
         if category_id.blank? && tags.blank? && canned_reply.blank? && hide_topic.blank? &&
              flag_post.blank?
@@ -42,6 +43,7 @@ def self.handle(
             temperature: 0,
             max_tokens: 700, # ~500 words
             user: Discourse.system_user,
+            stop_sequences: stop_sequences,
             feature_name: "llm_triage",
             feature_context: {
               automation_id: automation&.id,

lib/completions/endpoints/canned_response.rb

@@ -17,7 +17,7 @@ def normalize_model_params(model_params)
           model_params
         end
 
-        attr_reader :responses, :completions, :dialect
+        attr_reader :responses, :completions, :dialect, :model_params
 
         def prompt_messages
           dialect.prompt.messages
@@ -26,12 +26,13 @@ def prompt_messages
         def perform_completion!(
           dialect,
           _user,
-          _model_params,
+          model_params,
           feature_name: nil,
           feature_context: nil,
           partial_tool_calls: false
         )
           @dialect = dialect
+          @model_params = model_params
           response = responses[completions]
           if response.nil?
             raise CANNED_RESPONSE_ERROR,

spec/lib/modules/automation/llm_triage_spec.rb

@@ -180,4 +180,22 @@ def triage(**args)
       expect(triage_prompt.messages.last[:upload_ids]).to contain_exactly(post_upload.id)
     end
   end
+
+  it "includes stop_sequences in the completion call" do
+    sequences = %w[GOOD BAD]
+
+    DiscourseAi::Completions::Llm.with_prepared_responses(["bad"]) do |spy|
+      triage(
+        post: post,
+        model: "custom:#{llm_model.id}",
+        system_prompt: "test %%POST%%",
+        search_for_text: "bad",
+        flag_post: true,
+        automation: nil,
+        stop_sequences: sequences,
+      )
+
+      expect(spy.model_params[:stop_sequences]).to contain_exactly(*sequences)
+    end
+  end
 end