[dataset] name = "cohere" size = "1m" table = "cohere_wiki" expected_rows = 1000000 dims = 1024 vector_field = "emb" s3_base_path = "s3://paradedb-benchmark/datasets/cohere/v1" # Both sections match the pdbench Compose service with CI's CPU/worker override. # `pdbench search` checks [postgres.search] before a sweep. CI builds the index with # pdbench-load, so [postgres.build] is not checked; the profile format requires it. [postgres.build] shared_buffers = "8GB" maintenance_work_mem = "8GB" max_parallel_maintenance_workers = 7 [postgres.search] shared_buffers = "8GB" max_parallel_workers = 8 max_worker_processes = 32 [index] target_segment_count = 8 vector_router_quantization = true [search] runs = 1 warmup_runs = 1 explain_sample = 100 [search.session] "cohere.titles_1pct" = "battle" "cohere.titles_10pct" = "year" max_parallel_workers_per_gather = "4" jit = "off" plan_cache_mode = "force_custom_plan" [search.sweep] param = "paradedb.vector_cluster_max_probe" min = 1.0 max = 100_000.0 start = 100.0 integer = true targets = [0.9, 0.95, 0.99] tolerance = 0.000001 max_probes = 20