change the meta prompt and add the number of the expected responses t…

…o it, and adapt the all functions to this change.

src/functions/aggregate/llm_first_or_last/implementation.cpp

@@ -31,6 +31,8 @@ nlohmann::json LlmFirstOrLast::Evaluate(nlohmann::json& tuples) {
 
     do {
         accumulated_tuples_tokens = Tiktoken::GetNumTokens(batch_tuples.dump());
+        accumulated_tuples_tokens +=
+            Tiktoken::GetNumTokens(PromptManager::ConstructNumTuples(static_cast<int>(tuples.size())));
         accumulated_tuples_tokens +=
             Tiktoken::GetNumTokens(PromptManager::ConstructInputTuplesHeader(tuples[start_index]));
         while (accumulated_tuples_tokens < static_cast<unsigned int>(available_tokens) &&

src/functions/aggregate/llm_reduce/implementation.cpp

@@ -32,6 +32,8 @@ nlohmann::json LlmReduce::ReduceLoop(const std::vector<nlohmann::json>& tuples,
 
     do {
         accumulated_tuples_tokens = Tiktoken::GetNumTokens(batch_tuples.dump());
+        accumulated_tuples_tokens +=
+            Tiktoken::GetNumTokens(PromptManager::ConstructNumTuples(static_cast<int>(tuples.size())));
         accumulated_tuples_tokens +=
             Tiktoken::GetNumTokens(PromptManager::ConstructInputTuplesHeader(tuples[start_index]));
         while (accumulated_tuples_tokens < static_cast<unsigned int>(available_tokens) &&

src/functions/aggregate/llm_rerank/implementation.cpp

@@ -40,6 +40,8 @@ nlohmann::json LlmRerank::SlidingWindow(nlohmann::json& tuples) {
         next_tuples.clear();
         batch_size = half_batch;
         accumulated_rows_tokens = Tiktoken::GetNumTokens(window_tuples.dump());
+        accumulated_rows_tokens +=
+            Tiktoken::GetNumTokens(PromptManager::ConstructNumTuples(static_cast<int>(tuples.size())));
         accumulated_rows_tokens +=
             Tiktoken::GetNumTokens(PromptManager::ConstructInputTuplesHeader(tuples[start_index]));
         while (available_tokens - accumulated_rows_tokens > 0 && start_index >= 0) {

src/functions/scalar/scalar.cpp

@@ -32,6 +32,8 @@ nlohmann::json ScalarFunctionBase::BatchAndComplete(const std::vector<nlohmann::
         int start_index = 0;
 
         do {
+            accumulated_tuples_tokens +=
+                Tiktoken::GetNumTokens(PromptManager::ConstructNumTuples(static_cast<int>(tuples.size())));
             accumulated_tuples_tokens +=
                 Tiktoken::GetNumTokens(PromptManager::ConstructInputTuplesHeader(tuples[start_index]));
             while (accumulated_tuples_tokens < static_cast<unsigned int>(available_tokens) &&

src/include/flockmtl/prompt_manager/prompt_manager.hpp

@@ -33,16 +33,18 @@ class PromptManager {
 
     static PromptDetails CreatePromptDetails(const nlohmann::json& prompt_details_json);
 
-    static std::string ConstructMarkdownHeader(const nlohmann::json& tuple);
+    static std::string ConstructNumTuples(int num_tuples);
 
-    static std::string ConstructMarkdownSingleTuple(const nlohmann::json& tuple);
+    static std::string ConstructInputTuplesHeader(const nlohmann::json& tuple);
 
-    static std::string ConstructMarkdownArrayTuples(const nlohmann::json& tuples);
+    static std::string ConstructSingleInputTuple(const nlohmann::json& tuple);
+
+    static std::string ConstructInputTuples(const nlohmann::json& tuples);
 
     template <typename FunctionType>
     static std::string Render(const std::string& user_prompt, const nlohmann::json& tuples, FunctionType option) {
         auto prompt = PromptManager::GetTemplate(option);
-        auto markdown_tuples = PromptManager::ConstructMarkdownArrayTuples(tuples);
+        auto markdown_tuples = PromptManager::ConstructInputTuples(tuples);
         prompt = PromptManager::ReplaceSection(prompt, PromptSection::USER_PROMPT, user_prompt);
         prompt = PromptManager::ReplaceSection(prompt, PromptSection::TUPLES, markdown_tuples);
         return prompt;

src/include/flockmtl/prompt_manager/repository.hpp

@@ -11,15 +11,16 @@ enum class AggregateFunctionType { REDUCE, REDUCE_JSON, FIRST, LAST, RERANK };
 enum class ScalarFunctionType { COMPLETE_JSON, COMPLETE, FILTER };
 
 constexpr auto META_PROMPT =
-    "You are a semantic analysis tool for DBMS. The tool will analyze each tuple in the provided data and respond to "
-    "user requests based on this context.\n\nUser Prompt:\n\n- {{USER_PROMPT}}\n\nTuples "
+    "You are FlockMTL a semantic analysis tool for DBMS. You will analyze each tuple in the provided data and respond "
+    "to "
+    "the user prompt.\n\nUser Prompt:\n\n- {{USER_PROMPT}}\n\nTuples "
     "Table:\n\n{{TUPLES}}\n\nInstructions:\n\n{{INSTRUCTIONS}}\n\nExpected Response Format:\n\n{{RESPONSE_FORMAT}}";
 
 class INSTRUCTIONS {
 public:
     static constexpr auto SCALAR_FUNCTION =
         "- The response should be directly relevant to each tuple without additional formatting, purely answering the "
-        "prompt as if each tuple were a standalone entity.\n- Use clear, context-relevant language to generate a "
+        "user prompt as if each tuple were a standalone entity.\n- Use clear, context-relevant language to generate a "
         "meaningful and concise answer for each tuple.";
     static constexpr auto AGGREGATE_FUNCTION =
         "- For each tuple in the provided data, evaluate the relevant attribute(s) based on the user prompt.\n- After "
@@ -36,18 +37,18 @@ class RESPONSE_FORMAT {
 public:
     // Scalar Functions
     static constexpr auto COMPLETE_JSON =
-        "The system should interpret database tuples and provide a response to the user's prompt for each tuple in a "
-        "JSON format that contains the necessary columns for the answer.\n\nThe tool should respond in JSON format as "
-        "follows:\n\n```json\n{\t\"tuples\": [\n\t\t{<response 1>},\n\t\t{<response 2>},\n\t\t...\n\t\t{<response "
-        "n>}\n\t]\n}\n```";
+        "You should return the responses to the user's prompt for each tuple in a "
+        "JSON format that contains the necessary columns for the answer.\n\nThe tool should respond in JSON format:\n\n"
+        "```json\n{\"tuples\": [{<response>},{<response>}, ..., {<response>}]}\n```";
     static constexpr auto COMPLETE =
-        "The system should interpret database tuples and provide a response to the user's prompt for each tuple in "
-        "plain text.\n\tThe tool should respond in JSON format as follows:\n\n```json\n{\"tuples\": [\"<response 1>\", "
-        "\"<response 2>\", ... , \"<response n>\"]}";
+        "You should return the responses to the user's prompt for each tuple in plain text. Ensure no tuple is "
+        "missed.\n"
+        "Respond in the following JSON format:\n\n"
+        "```json\n{\"tuples\": [\"<response>\", \"<response>\", ..., \"<response>\"]}\n```";
     static constexpr auto FILTER =
-        "The system should interpret database tuples and provide a response to the user's prompt for each tuple in a "
+        "You should return the responses to the user's prompt for each tuple in a "
         "BOOL format that would be true/false.\n\tThe tool should respond in JSON format as "
-        "follows:\n\n```json\n{\"tuples\": [<bool response 1>, <bool response 2>, ... , <bool response n>]}";
+        "follows:\n\n```json\n{\"tuples\": [<bool_response>, <bool_response>, ... , <bool_response>]}\n```";
 
     // Aggregate Functions
     static constexpr auto REDUCE =

src/prompt_manager/prompt_manager.cpp

@@ -56,8 +56,13 @@ std::string PromptManager::ConstructSingleInputTuple(const nlohmann::json& tuple
     return tuple_str;
 }
 
+std::string PromptManager::ConstructNumTuples(const int num_tuples) {
+    return "- The Number of Tuples to Generate Responses for: " + std::to_string(num_tuples) + "\n\n";
+}
+
 std::string PromptManager::ConstructInputTuples(const nlohmann::json& tuples) {
     auto tuples_str = std::string("");
+    tuples_str += PromptManager::ConstructNumTuples(static_cast<int>(tuples.size()));
     tuples_str += PromptManager::ConstructInputTuplesHeader(tuples[0]);
     for (const auto& tuple : tuples) {
         tuples_str += PromptManager::ConstructSingleInputTuple(tuple);