Fixed scalar function to handle over/under predictions and updated ll…

…m_filter to handle null values in under predictions.

src/functions/scalar/llm_filter/implementation.cpp

@@ -36,6 +36,10 @@ std::vector<std::string> LlmFilter::Operation(duckdb::DataChunk& args) {
     std::vector<std::string> results;
     results.reserve(responses.size());
     for (const auto& response : responses) {
+        if (response.is_null()) {
+            results.emplace_back("True");
+            continue;
+        }
         results.push_back(response.dump());
     }
 

src/functions/scalar/scalar.cpp

@@ -53,12 +53,25 @@ nlohmann::json ScalarFunctionBase::BatchAndComplete(const std::vector<nlohmann::
                 response = Complete(batch_tuples, user_prompt, function_type, model);
             } catch (const ExceededMaxOutputTokensError&) {
                 batch_tuples.clear();
-                const auto new_batch_size = static_cast<int>(batch_size * 0.1);
+                const auto new_batch_size = static_cast<int>(batch_size / 10);
                 batch_size = batch_size == 1 ? new_batch_size == 0 : new_batch_size;
                 accumulated_tuples_tokens = 0;
                 start_index = 0;
                 continue;
             }
+
+            if (response.size() < batch_tuples.size()) {
+                for (int i = static_cast<int>(response.size()); i < batch_tuples.size(); i++) {
+                    response.push_back(nullptr);
+                }
+            } else if (response.size() > batch_size) {
+                auto new_response = nlohmann::json::array();
+                for (int i = 0; i < batch_size; i++) {
+                    new_response.push_back(response[i]);
+                }
+                response = new_response;
+            }
+
             auto output_tokens_per_tuple = Tiktoken::GetNumTokens(response.dump()) / batch_tuples.size();
 
             batch_size = model.GetModelDetails().max_output_tokens / output_tokens_per_tuple;