diff --git a/src/09python_ffi/src/main.cpp b/src/09python_ffi/src/main.cpp
index 48a4ea6f..0cf0323c 100644
--- a/src/09python_ffi/src/main.cpp
+++ b/src/09python_ffi/src/main.cpp
@@ -3,6 +3,7 @@
 #include "import.h"
 #include "llm/operators.h"
 #include "onnx/operators.h"
+#include "search.h"
 #include <pybind11/stl.h>// keep this line to convert stl types
 
 namespace py = pybind11;
@@ -26,6 +27,7 @@ namespace refactor::python_ffi {
         py::class_<Device::Blob, Arc<Device::Blob>>(m, "Pinned"      );
 
         m   .def("config_log"      , &configLog                  , return_::automatic )
+            .def("random_search"   , &randomSearch               , return_::move      )
             .def("find_device"     , &findDevice                 , return_::move      )
             .def("_make_operator"  , &makeOp                     , return_::move      )
             .def("_make_tensor"    , &makeTensor                 , return_::move      )
diff --git a/src/09python_ffi/src/search.cpp b/src/09python_ffi/src/search.cpp
new file mode 100644
index 00000000..63961081
--- /dev/null
+++ b/src/09python_ffi/src/search.cpp
@@ -0,0 +1,69 @@
+#include "search.h"
+#include "functions.h"
+#include <random>
+
+namespace refactor::python_ffi {
+
+    pybind11::array randomSearch(pybind11::array logits_, int topK, float topP, float temperature) {
+        auto shape = std::span(logits_.shape(), logits_.ndim());
+        ASSERT(!shape.empty(), "");
+        auto shapeBack = shape.begin() + shape.size() - 1;
+        auto batch = std::accumulate(shape.begin(), shapeBack, 1l, std::multiplies()),
+             vocabSize = *shapeBack;
+
+        std::random_device rd;
+        std::mt19937 gen(rd());
+        std::uniform_real_distribution<float> rng(0, 1);
+        std::vector<int64_t> result(batch);
+        for (auto i : range0_(batch)) {
+            // cast
+            std::vector<float> logits(vocabSize);
+            if (auto type = parseNumpyDType(logits_.dtype()); type == DataType::FP16) {
+                auto data = reinterpret_cast<fp16_t const *>(logits_.data()) + i * vocabSize;
+                std::transform(data, data + logits.size(), logits.begin(), [=](auto x) { return x.to_f32() / temperature; });
+            } else if (type == DataType::F32) {
+                auto data = reinterpret_cast<float const *>(logits_.data()) + i * vocabSize;
+                std::transform(data, data + logits.size(), logits.begin(), [=](auto x) { return x / temperature; });
+            } else {
+                RUNTIME_ERROR("unsupported data type.");
+            }
+            std::vector<std::pair<float, int>> probabilities(vocabSize);
+            // softmax
+            for (auto max = *std::max_element(logits.begin(), logits.end());
+                 auto j : range0_(vocabSize)) {
+                auto p = std::exp(logits[j] - max);
+                probabilities[j] = {p, j};
+            }
+            auto k = 0;
+            {// topK + topP
+                std::sort(probabilities.begin(), probabilities.end(),
+                          [](auto a, auto b) { return a.first > b.first; });
+                for (auto cum = 0.f; auto i : range0_(topK)) {
+                    ++k;
+                    if ((cum += probabilities[i].first) > topP) {
+                        break;
+                    }
+                }
+            }
+            auto chosen = false;
+            auto p = rng(gen);
+            // re-softmax
+            for (auto max = probabilities[0].first;
+                 auto j : range0_(k)) {
+                auto p_ = std::exp(probabilities[j].first - max);
+                if (p_ >= p) {
+                    result[i] = probabilities[j].second;
+                    chosen = true;
+                    break;
+                }
+                p -= p_;
+            }
+            if (!chosen) {
+                result[i] = probabilities[k - 1].second;
+            }
+        }
+
+        return pybind11::array(buildNumpyDType(DataType::I64), std::span(shape.begin(), shape.size() - 1), result.data());
+    }
+
+}// namespace refactor::python_ffi
diff --git a/src/09python_ffi/src/search.h b/src/09python_ffi/src/search.h
new file mode 100644
index 00000000..f9d8d398
--- /dev/null
+++ b/src/09python_ffi/src/search.h
@@ -0,0 +1,12 @@
+#ifndef PYTHON_FFI_SEARCH_H
+#define PYTHON_FFI_SEARCH_H
+
+#include <pybind11/numpy.h>
+
+namespace refactor::python_ffi {
+
+    pybind11::array randomSearch(pybind11::array, int topK, float topP, float temperature);
+
+}// namespace refactor::python_ffi
+
+#endif// PYTHON_FFI_SEARCH_H