{
  "measured_at": "2026-10-09 08:08",
  "machine": {
    "gpu": "NVIDIA GeForce RTX 5060 Laptop GPU",
    "vram_gb": 8.0,
    "cpu": "Intel64 Family 6 Model 183 Stepping 1, GenuineIntel",
    "os": "Windows-11",
    "torch": "2.14.0+cu130"
  },
  "model": {
    "name": "EmbeddingGemma2 (text encoder only)",
    "params_m": 271,
    "dtype": "bfloat16",
    "dim": 768,
    "max_seq_len": 8192,
    "device": "cuda:0"
  },
  "cold_start_ms": {
    "imports": 7184.5,
    "model_load": 2560.7,
    "first_encode": 572.0,
    "first_query_graph_capture": 316.9
  },
  "cuda_graphs": true,
  "repos": {
    "codesearch": {
      "files": 21,
      "chunks": 70,
      "doc_tokens": 39488,
      "full_index_s": 2.29,
      "index_tokens_per_s": 17245,
      "noop_refresh_ms": 5.0,
      "query_ms": {
        "p50": 5.4,
        "p95": 5.6,
        "embed_p50": 5.3
      },
      "peak_vram_gb": 2.45,
      "context": {
        "queries": [
          {
            "query": "where outbound network connections are blocked",
            "hit_files": 4,
            "tokens_returned": 2926,
            "tokens_whole_files": 12387
          },
          {
            "query": "how source files are split into chunks",
            "hit_files": 4,
            "tokens_returned": 2704,
            "tokens_whole_files": 9791
          },
          {
            "query": "batching texts under a token budget",
            "hit_files": 4,
            "tokens_returned": 2702,
            "tokens_whole_files": 6740
          },
          {
            "query": "incremental reindex of changed files",
            "hit_files": 6,
            "tokens_returned": 2542,
            "tokens_whole_files": 18056
          },
          {
            "query": "cosine similarity search over stored embeddings",
            "hit_files": 5,
            "tokens_returned": 2722,
            "tokens_whole_files": 10598
          },
          {
            "query": "loading the model on gpu in bfloat16",
            "hit_files": 5,
            "tokens_returned": 2524,
            "tokens_whole_files": 7706
          }
        ],
        "tokens_returned_avg": 2687,
        "tokens_whole_files_avg": 10880
      }
    },
    "caliper": {
      "files": 97,
      "chunks": 1219,
      "doc_tokens": 804403,
      "full_index_s": 58.93,
      "index_tokens_per_s": 13649,
      "noop_refresh_ms": 42.7,
      "query_ms": {
        "p50": 6.3,
        "p95": 8.6,
        "embed_p50": 6.1
      },
      "peak_vram_gb": 2.49,
      "context": {
        "queries": [
          {
            "query": "where the hero point cloud is rendered",
            "hit_files": 6,
            "tokens_returned": 6648,
            "tokens_whole_files": 213756
          },
          {
            "query": "sign-in gate that locks the tools",
            "hit_files": 7,
            "tokens_returned": 3748,
            "tokens_whole_files": 142445
          },
          {
            "query": "how provider api keys are redacted from logs",
            "hit_files": 5,
            "tokens_returned": 3580,
            "tokens_whole_files": 41625
          },
          {
            "query": "scoring math for a use case",
            "hit_files": 4,
            "tokens_returned": 3877,
            "tokens_whole_files": 47809
          },
          {
            "query": "slider tick marks drawn from data attributes",
            "hit_files": 5,
            "tokens_returned": 3826,
            "tokens_whole_files": 245046
          },
          {
            "query": "user supplied url checked before requests",
            "hit_files": 7,
            "tokens_returned": 2553,
            "tokens_whole_files": 23373
          }
        ],
        "tokens_returned_avg": 4039,
        "tokens_whole_files_avg": 119009
      }
    }
  }
}