Move specialized LLM skills out of the system prompt.

profiler/CMakeLists.txt

@@ -144,6 +144,8 @@ set(PROFILER_FILES
 )
 
 Embed(PROFILER_FILES SystemPrompt src/llm/system.prompt.md)
+Embed(PROFILER_FILES SkillCallstack src/llm/skill.callstack.md)
+Embed(PROFILER_FILES SkillOptimization src/llm/skill.optimization.md)
 Embed(PROFILER_FILES ToolsJson src/llm/tools.json)
 Embed(PROFILER_FILES FontFixed src/font/FiraCode-Retina.ttf)
 Embed(PROFILER_FILES FontIcons src/font/Font\ Awesome\ 6\ Free-Solid-900.otf)

profiler/src/llm/skill.callstack.md

@@ -0,0 +1,6 @@
+### Inspecting callstacks
+
+1. Focus on user's code. Ignore standard library boilerplate.
+2. Retrieve source code to verify callstack validity. Source locations in callstacks are return locations, and the call site may actually be near the reported source line.
+3. Top of the callstack is the most interesting, as it shows what the program is doing *now*. The bottom of the callstack shows what the program did to do what it's doing.
+4. If the callstack contains Tracy's crash handler, the profiled program has crashed. In this case, ignore the crash handler and any functions it may be calling. The crash happened *before* the handler intercepted it.

profiler/src/llm/skill.optimization.md

@@ -0,0 +1,8 @@
+### Program optimization
+
+1. Start by mapping the assembly instructions to the source code. All the reasoning should be performed with source code first. The assembly can only be used as a supplementary source.
+2. Analyze the available data, looking where the majority of the run time is spent. Always look at the code as a whole. Do not stop after finding a bunch of interesting spots.
+3. Figure out what algorithms are in use, how the data is structured and how it flows, reason about trade-offs taken.
+4. Reason if the code can be made to perform better. Note that some code will already be optimal, despite having hot spots.
+5. Formulate the optimization strategies and present them to the user.
+6. Do not provide concrete speed up percentages. It is only possible to know how faster the code is by measuring it after the changes. You can't do that.

profiler/src/llm/system.prompt.md

@@ -53,20 +53,8 @@ Insert links inline in the text, for example: "Function xyz() is located at [lin
 
 ## Case specific operation
 
-In certain situations you must use a specialized workflow.
+Specialized instructions and workflows for specific tasks are provided with the `skill` tool. If the task description matches the skill description, you must load the skill in question to gather the required abilities, *before* doing anything else.
 
-### Program optimization
+Available skills:
 
-1. Start by mapping the assembly instructions to the source code. All the reasoning should be performed with source code first. The assembly can only be used as a supplementary source.
-2. Analyze the available data, looking where the majority of the run time is spent. Always look at the code as a whole. Do not stop after finding a bunch of interesting spots.
-3. Figure out what algorithms are in use, how the data is structured and how it flows, reason about trade-offs taken.
-4. Reason if the code can be made to perform better. Note that some code will already be optimal, despite having hot spots.
-5. Formulate the optimization strategies and present them to the user.
-6. Do not provide concrete speed up percentages. It is only possible to know how faster the code is by measuring it after the changes. You can't do that.
-
-### Inspecting callstacks
-
-1. Focus on user's code. Ignore standard library boilerplate.
-2. Retrieve source code to verify callstack validity. Source locations in callstacks are return locations, and the call site may actually be near the reported source line.
-3. Top of the callstack is the most interesting, as it shows what the program is doing *now*. The bottom of the callstack shows what the program did to do what it's doing.
-4. If the callstack contains Tracy's crash handler, the profiled program has crashed. In this case, ignore the crash handler and any functions it may be calling. The crash happened *before* the handler intercepted it.
+%SKILLS%

profiler/src/llm/tools.json

@@ -1,4 +1,21 @@
 [
+{
+  "type": "function",
+  "function": {
+    "name": "skill",
+    "description": "Learn a specialized skill for the task needing specific knowledge.",
+    "parameters": {
+      "type": "object",
+      "properties": {
+        "name": {
+          "type": "string",
+          "description": "Name of the skill, chosen from the list of available skills."
+        }
+      },
+      "required": ["name"]
+    }
+  }
+},
 {
   "type": "function",
   "function": {

profiler/src/profiler/TracyLlm.cpp

@@ -17,6 +17,8 @@
 #include "../public/common/TracySystem.hpp"
 
 #include "data/SystemPrompt.hpp"
+#include "data/SkillCallstack.hpp"
+#include "data/SkillOptimization.hpp"
 #include "data/ToolsJson.hpp"
 
 namespace tracy
@@ -42,6 +44,9 @@ TracyLlm::TracyLlm( Worker& worker, View& view, const TracyManualData& manual )
         atexit( curl_global_cleanup );
     }
 
+    AddSkill( "callstack", "Analyze content of a call stack or crash trace", Unembed( SkillCallstack ) );
+    AddSkill( "optimization", "General code optimization workflow", Unembed( SkillOptimization ) );
+
     m_systemPrompt = Unembed( SystemPrompt );
     auto toolsJson = Unembed( ToolsJson );
     m_toolsJson = nlohmann::json::parse( toolsJson->data(), toolsJson->data() + toolsJson->size() );
@@ -51,8 +56,8 @@ TracyLlm::TracyLlm( Worker& worker, View& view, const TracyManualData& manual )
     ResetChat();
 
     m_api = std::make_unique<TracyLlmApi>();
-    m_chatUi = std::make_unique<TracyLlmChat>( view, worker );
-    m_tools = std::make_unique<TracyLlmTools>( worker, manual );
+    m_chatUi = std::make_unique<TracyLlmChat>( view, worker, m_skills );
+    m_tools = std::make_unique<TracyLlmTools>( worker, manual, m_skills );
 
     m_busy = true;
     QueueConnect();
@@ -872,15 +877,20 @@ void TracyLlm::UpdateSystemPrompt()
     static constexpr std::string_view UserToken = "%USER%";
     static constexpr std::string_view TimeToken = "%TIME%";
     static constexpr std::string_view ProgramNameToken = "%PROGRAMNAME%";
+    static constexpr std::string_view SkillsToken = "%SKILLS%";
 
     auto userName = GetUserFullName();
     if( !userName ) userName = GetUserLogin();
 
+    std::string skills;
+    for( auto& skill : m_skills ) skills += skill.name + ": " + skill.description + "\n";
+
     auto systemPrompt = std::string( m_systemPrompt->data(), m_systemPrompt->size() );
 
     Replace( systemPrompt, UserToken, userName );
     Replace( systemPrompt, TimeToken, m_tools->GetCurrentTime() );
     Replace( systemPrompt, ProgramNameToken, m_worker.GetCaptureProgram() );
+    Replace( systemPrompt, SkillsToken, skills );
 
     if( !m_api )
     {
@@ -1304,4 +1314,9 @@ bool TracyLlm::OnResponse( const nlohmann::json& json )
     return true;
 }
 
+void TracyLlm::AddSkill( std::string&& name, std::string&& description, const std::shared_ptr<EmbedData>& content )
+{
+    m_skills.emplace_back( std::move( name ), std::move( description ), std::string( content->data(), content->size() ) );
+}
+
 }

profiler/src/profiler/TracyLlm.hpp

@@ -23,6 +23,13 @@ class TracyManualData;
 class View;
 class Worker;
 
+struct LlmSkill
+{
+    std::string name;
+    std::string description;
+    std::string content;
+};
+
 class TracyLlm
 {
     enum class Task
@@ -81,6 +88,8 @@ private:
     void AppendResponse( const char* name, const nlohmann::json& delta );
     bool OnResponse( const nlohmann::json& json );
 
+    void AddSkill( std::string&& name, std::string&& description, const std::shared_ptr<EmbedData>& content );
+
     std::unique_ptr<TracyLlmApi> m_api;
     std::unique_ptr<TracyLlmChat> m_chatUi;
     std::unique_ptr<TracyLlmTools> m_tools;
@@ -111,6 +120,7 @@ private:
     std::vector<nlohmann::json> m_chat;
     std::string m_summary;
 
+    std::vector<LlmSkill> m_skills;
     std::shared_ptr<EmbedData> m_systemPrompt;
     nlohmann::json m_toolsJson;
 

profiler/src/profiler/TracyLlmChat.cpp

@@ -5,6 +5,7 @@
 #include <string>
 
 #include "TracyImGui.hpp"
+#include "TracyLlm.hpp"
 #include "TracyLlmChat.hpp"
 #include "TracyMouse.hpp"
 #include "TracyPrint.hpp"
@@ -68,7 +69,7 @@ static const char* utfendl( const char* str, int len )
 }
 
 
-static std::string ToolCallDescription( const nlohmann::json& json )
+std::string TracyLlmChat::ToolCallDescription( const nlohmann::json& json ) const
 {
     if( !json.contains( "arguments" ) ) return "";
     nlohmann::json args;
@@ -127,13 +128,22 @@ static std::string ToolCallDescription( const nlohmann::json& json )
         if( args.contains( "path" ) ) path = ", path: " + args["path"].get_ref<const std::string&>();
         return "Source search: " + args["query"].get_ref<const std::string&>() + caseInsensitive + path;
     }
+    else if( name == "skill" )
+    {
+        if( !args.contains( "name" ) ) return "";
+        auto skill = args["name"].get_ref<const std::string&>();
+        auto it = std::ranges::find_if( m_skills, [&skill]( const auto& s ) { return s.name == skill; } );
+        if( it == m_skills.end() ) return "";
+        return "Learn skill: " + it->description;
+    }
     return "";
 }
 
 
-TracyLlmChat::TracyLlmChat( View& view, Worker& worker )
+TracyLlmChat::TracyLlmChat( View& view, Worker& worker, const std::vector<LlmSkill>& skills )
     : m_width( new float[NumRoles] )
     , m_markdown( &view, &worker )
+    , m_skills( skills )
 {
 }
 

profiler/src/profiler/TracyLlmChat.hpp

@@ -2,12 +2,14 @@
 #define __TRACYLLMCHAT_HPP__
 
 #include <nlohmann/json.hpp>
+#include <vector>
 
 #include "TracyMarkdown.hpp"
 
 namespace tracy
 {
 
+struct LlmSkill;
 class View;
 class Worker;
 
@@ -35,7 +37,7 @@ public:
         ToolCall
     };
 
-    TracyLlmChat( View& view, Worker& worker );
+    TracyLlmChat( View& view, Worker& worker, const std::vector<LlmSkill>& skills );
     ~TracyLlmChat();
 
     void Begin();
@@ -50,6 +52,8 @@ private:
 
     void PrintThink( const char* str, size_t size );
 
+    [[nodiscard]] std::string ToolCallDescription( const nlohmann::json& json ) const;
+
     float* m_width;
     float m_maxWidth;
 
@@ -61,6 +65,8 @@ private:
 
     Markdown m_markdown;
     std::string m_label;
+
+    const std::vector<LlmSkill>& m_skills;
 };
 
 }

profiler/src/profiler/TracyLlmTools.cpp

@@ -10,6 +10,7 @@
 #include <regex>
 
 #include "TracyConfig.hpp"
+#include "TracyLlm.hpp"
 #include "TracyLlmApi.hpp"
 #include "TracyLlmTools.hpp"
 #include "TracyManualData.hpp"
@@ -78,9 +79,10 @@ static std::unique_ptr<pugi::xml_document> ParseHtml( const std::string& html )
     return doc;
 }
 
-TracyLlmTools::TracyLlmTools( Worker& worker, const TracyManualData& manual )
+TracyLlmTools::TracyLlmTools( Worker& worker, const TracyManualData& manual, const std::vector<LlmSkill>& skills )
     : m_worker( worker )
     , m_manual( manual )
+    , m_skills( skills )
 {
     int idx = 0;
     for( auto& chunk : m_manual.GetChunks() )
@@ -182,6 +184,10 @@ std::string TracyLlmTools::HandleToolCalls( const std::string& tool, const nlohm
             std::string empty;
             return SourceSearch( Param( "query" ), ParamOptBool( "case_insensitive", false ), ParamOptString( "path", empty ) );
         }
+        else if( tool == "skill" )
+        {
+            return GetSkill( Param( "name" ) );
+        }
         return "Unknown tool call: " + tool;
     }
     catch( const std::exception& e )
@@ -1008,4 +1014,11 @@ std::string TracyLlmTools::SourceSearch( std::string query, bool caseInsensitive
     return ret;
 }
 
+std::string TracyLlmTools::GetSkill( const std::string& name ) const
+{
+    auto it = std::ranges::find_if( m_skills, [&name]( const auto& skill ) { return skill.name == name; } );
+    if( it == m_skills.end() ) return "No such skill.";
+    return it->content;
+}
+
 }

profiler/src/profiler/TracyLlmTools.hpp

@@ -16,6 +16,7 @@ class EmbedData;
 namespace tracy
 {
 
+struct LlmSkill;
 class TracyLlmApi;
 class TracyManualData;
 class Worker;
@@ -31,7 +32,7 @@ public:
         float progress = 0;
     };
 
-    TracyLlmTools( Worker& worker, const TracyManualData& manual );
+    TracyLlmTools( Worker& worker, const TracyManualData& manual, const std::vector<LlmSkill>& skills );
     ~TracyLlmTools();
 
     std::string HandleToolCalls( const std::string& tool, const nlohmann::json& json, TracyLlmApi& api, int contextSize, bool hasEmbeddingsModel );
@@ -62,6 +63,7 @@ private:
     std::string SearchManual( const std::string& query, TracyLlmApi& api, bool hasEmbeddingsModel );
     std::string SourceFile( const std::string& file, uint32_t line, uint32_t context, uint32_t contextBack ) const;
     std::string SourceSearch( std::string query, bool caseInsensitive, const std::string& path ) const;
+    std::string GetSkill( const std::string& name ) const;
 
     void ManualEmbeddingsWorker( TracyLlmApi& api );
 
@@ -79,6 +81,7 @@ private:
 
     Worker& m_worker;
     const TracyManualData& m_manual;
+    const std::vector<LlmSkill>& m_skills;
 };
 
 }