Figure Studio

Edits stay in this tab. Copy the Markdown into a note to publish; nothing is uploaded or saved here.

Existing measurements from the BCR note, with their source and experimental limits.

Markdown sourceEditable

```chart
{
  "type": "bar",
  "title": "Exact-prefix serving latency",
  "description": "BCR latency divided by RadixCache latency. Ratios above one mean BCR was slower.",
  "caption": "Qwen2.5-7B-Instruct on an A100 80GB; two cache-friendly workloads. A ratio of 1 is equal latency; values above 1 are slower. Reuse telemetry matched. These measurements do not show a serving-speed advantage for BCR.",
  "source": {
    "label": "BCR-memory-4 · Result 1, serving-fairness table",
    "url": "https://armanas.dev/artifact/1437e0df-444b-42f7-99cc-054c7dbf6605"
  },
  "xAxis": {
    "key": "workload",
    "label": "Workload",
    "type": "category"
  },
  "yAxis": {
    "key": "ratio",
    "label": "Latency / RadixCache",
    "unit": "×"
  },
  "series": [
    {
      "dataKey": "ttftP50",
      "name": "TTFT p50"
    },
    {
      "dataKey": "ttftP99",
      "name": "TTFT p99"
    },
    {
      "dataKey": "e2eP50",
      "name": "End-to-end p50"
    }
  ],
  "data": [
    {
      "workload": "system_repeat",
      "ttftP50": 1.081,
      "ttftP99": 1.197,
      "e2eP50": 1.003
    },
    {
      "workload": "rag_fanout",
      "ttftP50": 1.102,
      "ttftP99": 1.289,
      "e2eP50": 1.004
    }
  ]
}
```

Live preview

As it appears in a note

Fig. 1

Exact-prefix serving latencyQwen2.5-7B-Instruct on an A100 80GB; two cache-friendly workloads. A ratio of 1 is equal latency; values above 1 are slower. Reuse telemetry matched. These measurements do not show a serving-speed advantage for BCR.

BCR latency divided by RadixCache latency. Ratios above one mean BCR was slower.

Source: BCR-memory-4 · Result 1, serving-fairness table

View data (2 rows)

Plotted data: Exact-prefix serving latency. Missing observations are shown as a dash.
Workload	TTFT p50 (×)	TTFT p99 (×)	End-to-end p50 (×)
system_repeat	1.081	1.197	1.003
rag_fanout	1.102	1.289	1.004

Format guide

Edit the JSON inside the chart fence. Put observations in data; each series.dataKey names a numeric column. Set axis labels and units, then add a title, caption and source.

Choose line, bar, area or scatter. Use xAxis.type: "number" for numeric spacing. Straight segments are the default; use interpolation only when the method warrants it.

Switching examples keeps your edits in this page. Copy your work before leaving or reloading. Copying preserves the source and caption. To publish, paste the block into a note in the admin editor and use Preview before saving.

Workload

TTFT p50 (×)

TTFT p99 (×)

End-to-end p50 (×)

system_repeat

1.081

1.197

1.003

rag_fanout

1.102

1.289

1.004