[Uplift 1.62.x] AI Chat Mixtral fixes

components/ai_chat/core/browser/conversation_driver.cc

@@ -36,12 +36,19 @@ using ai_chat::mojom::CharacterType;
 using ai_chat::mojom::ConversationTurn;
 using ai_chat::mojom::ConversationTurnVisibility;
 
+namespace ai_chat {
+
 namespace {
+
 static const auto kAllowedSchemes = base::MakeFixedFlatSet<std::string_view>(
     {url::kHttpsScheme, url::kHttpScheme, url::kFileScheme, url::kDataScheme});
-}  // namespace
 
-namespace ai_chat {
+bool IsPremiumStatus(mojom::PremiumStatus status) {
+  return status == mojom::PremiumStatus::Active ||
+         status == mojom::PremiumStatus::ActiveDisconnected;
+}
+
+}  // namespace
 
 ConversationDriver::ConversationDriver(raw_ptr<PrefService> pref_service,
        raw_ptr<AIChatMetrics> ai_chat_metrics,
@@ -74,7 +81,7 @@ ConversationDriver::ConversationDriver(raw_ptr<PrefService> pref_service,
     credential_manager_->GetPremiumStatus(base::BindOnce(
         [](ConversationDriver* instance, mojom::PremiumStatus status) {
           instance->last_premium_status_ = status;
-          if (status == mojom::PremiumStatus::Inactive) {
+          if (!IsPremiumStatus(status)) {
             // Not premium
             return;
           }
@@ -721,9 +728,15 @@ void ConversationDriver::GetPremiumStatus(
 void ConversationDriver::OnPremiumStatusReceived(
     mojom::PageHandler::GetPremiumStatusCallback parent_callback,
     mojom::PremiumStatus premium_status) {
-  if (last_premium_status_ != premium_status &&
-      premium_status == mojom::PremiumStatus::Active) {
-    // Change model if we haven't already
+  // Maybe switch to premium model when user is newly premium and on a basic
+  // model
+  const bool should_switch_model =
+      // This isn't the first retrieval (that's handled in the constructor)
+      last_premium_status_ != mojom::PremiumStatus::Unknown &&
+      last_premium_status_ != premium_status &&
+      premium_status == mojom::PremiumStatus::Active &&
+      GetCurrentModel().access == mojom::ModelAccess::BASIC;
+  if (should_switch_model) {
     ChangeModel(features::kAIModelsPremiumDefaultKey.Get());
   }
   last_premium_status_ = premium_status;

components/ai_chat/core/browser/conversation_driver.h

@@ -161,7 +161,7 @@ class ConversationDriver {
   int64_t current_navigation_id_;
   bool is_same_document_navigation_ = false;
   mojom::APIError current_error_ = mojom::APIError::None;
-  mojom::PremiumStatus last_premium_status_ = mojom::PremiumStatus::Inactive;
+  mojom::PremiumStatus last_premium_status_ = mojom::PremiumStatus::Unknown;
 
   std::unique_ptr<mojom::ConversationTurn> pending_conversation_entry_;
   bool pending_message_needs_page_content_ = false;

components/ai_chat/core/browser/engine/engine_consumer_llama.cc

@@ -162,6 +162,7 @@ std::string BuildLlama2Prompt(
     const std::vector<ConversationTurn>& conversation_history,
     std::string page_content,
     const bool& is_video,
+    const bool& needs_general_seed,
     const std::string user_message) {
   // Always use a generic system message
   std::string system_message =
@@ -206,7 +207,9 @@ std::string BuildLlama2Prompt(
   if (conversation_history.empty() || conversation_history.size() <= 1) {
     return BuildLlama2FirstSequence(
         today_system_message, first_user_message, absl::nullopt,
-        l10n_util::GetStringUTF8(IDS_AI_CHAT_LLAMA2_GENERAL_SEED));
+        (needs_general_seed) ? std::optional(l10n_util::GetStringUTF8(
+                                   IDS_AI_CHAT_LLAMA2_GENERAL_SEED))
+                             : std::nullopt);
   }
 
   // Use the first two messages to build the first sequence,
@@ -227,7 +230,9 @@ std::string BuildLlama2Prompt(
   // Build the final subsequent exchange using the current turn.
   prompt += BuildLlama2SubsequentSequence(
       user_message, absl::nullopt,
-      l10n_util::GetStringUTF8(IDS_AI_CHAT_LLAMA2_GENERAL_SEED));
+      (needs_general_seed) ? std::optional(l10n_util::GetStringUTF8(
+                                 IDS_AI_CHAT_LLAMA2_GENERAL_SEED))
+                           : std::nullopt);
 
   // Trimming recommended by Meta
   // https://huggingface.co/meta-llama/Llama-2-13b-chat#intended-use
@@ -249,6 +254,8 @@ EngineConsumerLlamaRemote::EngineConsumerLlamaRemote(
   api_ = std::make_unique<RemoteCompletionClient>(
       model.name, stop_sequences, url_loader_factory, credential_manager);
 
+  needs_general_seed_ = base::StartsWith(model.name, "llama-2");
+
   max_page_content_length_ = model.max_page_content_length;
 }
 
@@ -332,8 +339,9 @@ void EngineConsumerLlamaRemote::GenerateAssistantResponse(
     GenerationCompletedCallback completed_callback) {
   const std::string& truncated_page_content =
       page_content.substr(0, max_page_content_length_);
-  std::string prompt = BuildLlama2Prompt(
-      conversation_history, truncated_page_content, is_video, human_input);
+  std::string prompt =
+      BuildLlama2Prompt(conversation_history, truncated_page_content, is_video,
+                        needs_general_seed_, human_input);
   DCHECK(api_);
   api_->QueryPrompt(prompt, {"</response>"}, std::move(completed_callback),
                     std::move(data_received_callback));

components/ai_chat/core/browser/engine/engine_consumer_llama.h

@@ -61,6 +61,7 @@ class EngineConsumerLlamaRemote : public EngineConsumer {
 
   std::unique_ptr<RemoteCompletionClient> api_ = nullptr;
 
+  bool needs_general_seed_ = true;
   int max_page_content_length_ = 0;
 
   base::WeakPtrFactory<EngineConsumerLlamaRemote> weak_ptr_factory_{this};

components/ai_chat/core/browser/models.cc

@@ -42,13 +42,13 @@ const std::vector<ai_chat::mojom::Model>& GetAllModels() {
   static const base::NoDestructor<std::vector<mojom::Model>> kModels({
       {"chat-leo-expanded", "mixtral-8x7b-instruct", "Mixtral", "Mistral AI",
        mojom::ModelEngineType::LLAMA_REMOTE, mojom::ModelCategory::CHAT,
-       kFreemiumAccess, 9000, 9700},
+       kFreemiumAccess, 8000, 9700},
       {"chat-claude-instant", "claude-instant-v1", "Claude Instant",
        "Anthropic", mojom::ModelEngineType::CLAUDE_REMOTE,
        mojom::ModelCategory::CHAT, kFreemiumAccess, 180000, 320000},
-      {"chat-basic", "llama-2-13b-chat", "llama2 13b", "Meta",
+      {"chat-basic", "llama-2-13b-chat", "Llama 2 13b", "Meta",
        mojom::ModelEngineType::LLAMA_REMOTE, mojom::ModelCategory::CHAT,
-       mojom::ModelAccess::BASIC, 9000, 9700},
+       mojom::ModelAccess::BASIC, 8000, 9700},
   });
   return *kModels;
 }

components/ai_chat/core/common/mojom/ai_chat.mojom

@@ -45,6 +45,7 @@ enum ModelAccess {
 };
 
 enum PremiumStatus {
+  Unknown,
   Inactive,
   Active,
   ActiveDisconnected,

components/resources/ai_chat_ui_strings.grdp

@@ -55,7 +55,7 @@
     Hi, I'm Leo. I'm a fully hosted AI assistant by Brave. I'm powered by Llama 13B, a model created by Meta to be performant and applicable to many use cases.
   </message>
   <message name="IDS_CHAT_UI_INTRO_MESSAGE_CHAT_LEO_EXPANDED" desc="AI Chat intro message for the expanded model">
-    Hi, I'm Leo. I'm a fully hosted AI assistant by Brave. I'm powered by Mixtral 7B, a model created by Mistral AI to handle advanced tasks.
+    Hi, I'm Leo. I'm a fully hosted AI assistant by Brave. I'm powered by Mixtral 8x7B, a model created by Mistral AI to handle advanced tasks.
   </message>
   <message name="IDS_CHAT_UI_INTRO_MESSAGE_CHAT_LEO_CLAUDE_INSTANT" desc="AI Chat intro message for the Claude Instant model">
     Hi, I'm Leo. I'm proxied by Brave and powered by Claude Instant, a model created by Anthropic to power conversational and text processing tasks.