From 381a2715c8046b47d58104d55450d15d864517ba Mon Sep 17 00:00:00 2001
From: Sam <sam.saffron@gmail.com>
Date: Sat, 1 Feb 2025 14:08:34 +1100
Subject: [PATCH] FEATURE: o3-mini supports (#1105)

1. Adds o3-mini presets
2. Adds support for reasoning effort
3. properly use "developer" messages for reasoning models
---
 app/models/llm_model.rb                       |  5 ++
 .../components/ai-llm-editor-form.gjs         | 47 ++++++++++++++-----
 config/locales/client.en.yml                  |  5 +-
 lib/completions/dialects/chat_gpt.rb          | 15 +++---
 lib/completions/endpoints/open_ai.rb          | 11 ++++-
 lib/completions/llm.rb                        |  6 ++-
 6 files changed, 66 insertions(+), 23 deletions(-)

diff --git a/app/models/llm_model.rb b/app/models/llm_model.rb
index f3244468..228c9335 100644
--- a/app/models/llm_model.rb
+++ b/app/models/llm_model.rb
@@ -34,6 +34,11 @@ class LlmModel < ActiveRecord::Base
         organization: :text,
         disable_native_tools: :checkbox,
         disable_streaming: :checkbox,
+        reasoning_effort: {
+          type: :enum,
+          values: %w[default low medium high],
+          default: "default",
+        },
       },
       mistral: {
         disable_native_tools: :checkbox,
diff --git a/assets/javascripts/discourse/components/ai-llm-editor-form.gjs b/assets/javascripts/discourse/components/ai-llm-editor-form.gjs
index dec5cb95..fb28de8a 100644
--- a/assets/javascripts/discourse/components/ai-llm-editor-form.gjs
+++ b/assets/javascripts/discourse/components/ai-llm-editor-form.gjs
@@ -114,10 +114,26 @@ export default class AiLlmEditorForm extends Component {
 
   @computed("args.model.provider")
   get metaProviderParams() {
-    return (
+    const params =
       this.args.llms.resultSetMeta.provider_params[this.args.model.provider] ||
-      {}
-    );
+      {};
+
+    return Object.entries(params).map(([field, value]) => {
+      if (typeof value === "string") {
+        return { field, type: value };
+      } else if (typeof value === "object") {
+        if (value.values) {
+          value = { ...value };
+          value.values = value.values.map((v) => {
+            return { id: v, name: v };
+          });
+        }
+        this.args.model.provider_params[field] =
+          this.args.model.provider_params[field] || value.default;
+        return { field, ...value };
+      }
+      return { field, type: "text" }; // fallback
+    });
   }
 
   @action
@@ -275,24 +291,31 @@ export default class AiLlmEditorForm extends Component {
             />
           </div>
         </div>
-        {{#each-in this.metaProviderParams as |field type|}}
-          <div class="control-group ai-llm-editor-provider-param__{{type}}">
+        {{#each this.metaProviderParams as |param|}}
+          <div
+            class="control-group ai-llm-editor-provider-param__{{param.type}}"
+          >
             <label>{{i18n
-                (concat "discourse_ai.llms.provider_fields." field)
+                (concat "discourse_ai.llms.provider_fields." param.field)
               }}</label>
-            {{#if (eq type "checkbox")}}
+            {{#if (eq param.type "enum")}}
+              <ComboBox
+                @value={{mut (get @model.provider_params param.field)}}
+                @content={{param.values}}
+              />
+            {{else if (eq param.type "checkbox")}}
               <Input
-                @type={{type}}
-                @checked={{mut (get @model.provider_params field)}}
+                @type={{param.type}}
+                @checked={{mut (get @model.provider_params param.field)}}
               />
             {{else}}
               <Input
-                @type={{type}}
-                @value={{mut (get @model.provider_params field)}}
+                @type={{param.type}}
+                @value={{mut (get @model.provider_params param.field)}}
               />
             {{/if}}
           </div>
-        {{/each-in}}
+        {{/each}}
         <div class="control-group">
           <label>{{i18n "discourse_ai.llms.tokenizer"}}</label>
           <ComboBox
diff --git a/config/locales/client.en.yml b/config/locales/client.en.yml
index bf76df2b..300d30c1 100644
--- a/config/locales/client.en.yml
+++ b/config/locales/client.en.yml
@@ -382,7 +382,9 @@ en:
           open_ai-gpt-4o: "High intelligence model for complex, multi-step tasks"
           open_ai-gpt-4o-mini: "Affordable and fast small model for lightweight tasks"
           open_ai-o1-mini: "Cost-efficient reasoning model"
-          open_ai-o1-preview: "Open AI's most capabale reasoning model"
+          open_ai-o1-preview: "Open AI's most capabale reasoning model (preview)"
+          open_ai-o1: "Open AI's most capable reasoning model"
+          open_ai-o3-mini: "Advanced Cost-efficient reasoning model"
           samba_nova-Meta-Llama-3-1-8B-Instruct: "Efficient lightweight multilingual model"
           samba_nova-Meta-Llama-3-1-70B-Instruct": "Powerful multipurpose model"
           mistral-mistral-large-latest: "Mistral's most powerful model"
@@ -439,6 +441,7 @@ en:
           provider_order: "Provider order (comma delimited list)"
           provider_quantizations: "Order of provider quantizations (comma delimited list eg: fp16,fp8)"
           disable_streaming: "Disable streaming completions (convert streaming to non streaming requests)"
+          reasoning_effort: "Reasoning effort (only applicable to reasoning models)"
 
       related_topics:
         title: "Related topics"
diff --git a/lib/completions/dialects/chat_gpt.rb b/lib/completions/dialects/chat_gpt.rb
index 9882adcf..33b00edb 100644
--- a/lib/completions/dialects/chat_gpt.rb
+++ b/lib/completions/dialects/chat_gpt.rb
@@ -40,11 +40,6 @@ module DiscourseAi
           llm_model.max_prompt_tokens - buffer
         end
 
-        # no support for streaming or tools or system messages
-        def is_gpt_o?
-          llm_model.provider == "open_ai" && llm_model.name.include?("o1-")
-        end
-
         def disable_native_tools?
           return @disable_native_tools if defined?(@disable_native_tools)
           !!@disable_native_tools = llm_model.lookup_custom_param("disable_native_tools")
@@ -60,14 +55,20 @@ module DiscourseAi
           end
         end
 
+        # developer messages are preferred on reasoning models
+        def supports_developer_messages?
+          llm_model.provider == "open_ai" &&
+            (llm_model.name.start_with?("o1") || llm_model.name.start_with?("o3"))
+        end
+
         def system_msg(msg)
           content = msg[:content]
           if disable_native_tools? && tools_dialect.instructions.present?
             content = content + "\n\n" + tools_dialect.instructions
           end
 
-          if is_gpt_o?
-            { role: "user", content: content }
+          if supports_developer_messages?
+            { role: "developer", content: content }
           else
             { role: "system", content: content }
           end
diff --git a/lib/completions/endpoints/open_ai.rb b/lib/completions/endpoints/open_ai.rb
index f1663449..f34f29db 100644
--- a/lib/completions/endpoints/open_ai.rb
+++ b/lib/completions/endpoints/open_ai.rb
@@ -43,7 +43,14 @@ module DiscourseAi
         private
 
         def disable_streaming?
-          @disable_streaming = llm_model.lookup_custom_param("disable_streaming")
+          @disable_streaming ||= llm_model.lookup_custom_param("disable_streaming")
+        end
+
+        def reasoning_effort
+          return @reasoning_effort if defined?(@reasoning_effort)
+          @reasoning_effort = llm_model.lookup_custom_param("reasoning_effort")
+          @reasoning_effort = nil if !%w[low medium high].include?(@reasoning_effort)
+          @reasoning_effort
         end
 
         def model_uri
@@ -60,6 +67,8 @@ module DiscourseAi
         def prepare_payload(prompt, model_params, dialect)
           payload = default_options.merge(model_params).merge(messages: prompt)
 
+          payload[:reasoning_effort] = reasoning_effort if reasoning_effort
+
           if @streaming_mode
             payload[:stream] = true
 
diff --git a/lib/completions/llm.rb b/lib/completions/llm.rb
index f7fe21a4..1db8f090 100644
--- a/lib/completions/llm.rb
+++ b/lib/completions/llm.rb
@@ -62,8 +62,10 @@ module DiscourseAi
                 {
                   id: "open_ai",
                   models: [
-                    { name: "o1-preview", tokens: 131_072, display_name: "o1" },
-                    { name: "o1-mini", tokens: 131_072, display_name: "o1 mini" },
+                    { name: "o3-mini", tokens: 200_000, display_name: "o3 Mini" },
+                    { name: "o1", tokens: 200_000, display_name: "o1" },
+                    { name: "o1-preview", tokens: 131_072, display_name: "o1 preview" },
+                    { name: "o1-mini", tokens: 131_072, display_name: "o1 Mini" },
                     { name: "gpt-4o", tokens: 131_072, display_name: "GPT-4 Omni" },
                     { name: "gpt-4o-mini", tokens: 131_072, display_name: "GPT-4 Omni Mini" },
                     { name: "gpt-4-turbo", tokens: 131_072, display_name: "GPT-4 Turbo" },