double free or corruption (out) during Search

I want to evaluate SPTAG on the streaming track of the big-ann-benchmarks. (https://github.com/harsha-simhadri/big-ann-benchmarks/blob/main/neurips23/README.md)
However, during the Search process, there was a double free or corruption (out) error. (It has to process 10K queries, but an error occurs in the middle)

Is there a problem with memory allocation and deallocation in the library? Or do I need to set some parameters?

Below is the benchmark code.

    def setup(self, dtype, max_pts, ndim):
        self.index = SPTAG.AnnIndex("BKT", "Float", ndim)

        self.index.SetBuildParam("NumberOfThreads", str(self.insert_threads), "Index")
        self.index.SetBuildParam("DistCalcMethod", self.translate_dist_fn(self._metric), "Index")
        self.max_pts = max_pts
        print('Index class constructed and ready for update/search')
    
    def insert(self, X, ids):
        self.index.SetBuildParam("NumberOfThreads", str(self.insert_threads), "Index")
        self.index.SetBuildParam("DistCalcMethod", self.translate_dist_fn(self._metric), "Index")
        p_meta = ''
        for i in ids:
            p_meta += str(i+1) + '\n'
        p_meta.encode()
        res = self.index.AddWithMetaData(X, p_meta, X.shape[0], True, False)

    def delete(self, ids):
        p_meta = ''
        for i in ids:
            p_meta += str(i+1) + '\n'
        p_meta.encode()

        self.index.DeleteByMetaData(p_meta)

    def query(self, X, k):
        result1 = []
        result2 = []
                
        for t in range(X.shape[0]):
            result = self.index.SearchWithMetaData(X[t], k)
            result1.append(result[2])
            result2.append(result[1])

        self.res = np.array(result1).reshape(-1, 10)
        self.query_dist = np.array(result2).reshape(-1, 10)

Start with the benchmark README and the setup, insert, delete, and query methods shown in the report. Reproduce the failure around SearchWithMetaData while processing 10K queries, then compare the AddWithMetaData and DeleteByMetaData calls with the search lifecycle. Done means identifying the cause of the memory error and confirming that the streaming benchmark completes without corruption.