{
  "id": "bbg-p0373-ai-models-and-llm-systems",
  "title": "Visual Token Context Compression",
  "chapter": "ai-models-and-llm-systems",
  "batch": "26",
  "rank": 251,
  "sourcePage": 373,
  "sourcePointer": "p. 373",
  "status": "accepted",
  "reviewerStatus": "reviewed",
  "fidelityScore": 0.9,
  "canvas": {
    "width": 960,
    "height": 640
  },
  "fireworksTechGraph": {
    "style": "style-1-flat-icon",
    "diagramType": "data-flow",
    "topologyNotes": [
      "source page render inspected",
      "extracted page text inspected",
      "source page render and extracted text inspected",
      "preserve DeepSeek-OCR idea: long context becomes image, compressed visual tokens feed the LLM, reducing attention cost",
      "omit source decorative art"
    ],
    "publicBoundary": [
      "original vector output",
      "no source pixels",
      "no source mark or long wording"
    ]
  },
  "callouts": [],
  "sourceReview": {
    "conceptAnchors": [
      "concept: long context window limit",
      "concept: convert context to image",
      "concept: visual tokens",
      "concept: lower attention cost",
      "concept: larger effective context"
    ],
    "labelSource": "curated",
    "semanticStatus": "reviewed"
  },
  "groups": [
    {
      "id": "compress",
      "label": "Context compression path",
      "x": 54,
      "y": 150,
      "w": 852,
      "h": 250
    }
  ],
  "shapes": [
    {
      "id": "context",
      "kind": "rect",
      "label": "Long context",
      "detail": "many tokens",
      "x": 88,
      "y": 228,
      "w": 132,
      "h": 58,
      "tone": "blue"
    },
    {
      "id": "image",
      "kind": "rect",
      "label": "Context image",
      "detail": "render",
      "x": 270,
      "y": 228,
      "w": 132,
      "h": 58,
      "tone": "orange"
    },
    {
      "id": "visual",
      "kind": "rect",
      "label": "Visual tokens",
      "detail": "compressed",
      "x": 452,
      "y": 228,
      "w": 132,
      "h": 58,
      "tone": "green"
    },
    {
      "id": "llm",
      "kind": "hex",
      "label": "LLM",
      "detail": "reason",
      "x": 636,
      "y": 216,
      "w": 132,
      "h": 82,
      "tone": "purple"
    },
    {
      "id": "cost",
      "kind": "rect",
      "label": "Lower cost",
      "detail": "less attention",
      "x": 452,
      "y": 342,
      "w": 132,
      "h": 54,
      "tone": "teal"
    }
  ],
  "connectors": [
    {
      "from": "context",
      "to": "image",
      "label": "render",
      "flow": "main"
    },
    {
      "from": "image",
      "to": "visual",
      "label": "compress",
      "flow": "main"
    },
    {
      "from": "visual",
      "to": "llm",
      "label": "feed",
      "flow": "data"
    },
    {
      "from": "visual",
      "to": "cost",
      "label": "reduce",
      "flow": "control",
      "dashed": true
    }
  ]
}