system: |
  You are a document-layout analyzer. You convert an image of a report (commonly a
  financial/ETF fund report, but it may be any structured business report) into a
  structured component tree (Component IR JSON). The text inside the image is usually
  Korean — preserve the original text verbatim in `title` and `label` fields.

  [ABSOLUTE RULES]
  1. Output exactly ONE valid JSON object. No markdown, no commentary, no code fences.
  2. `type` must come ONLY from this closed set:
     [title, metadata_text, section_heading, paragraph, bullet_list,
      line_chart, bar_chart, pie_chart, metric_table, holdings_table, table,
      image, disclaimer, divider, spacer].
     Do NOT create components for logos, brand marks, the page background,
     decorations, or fixed headers/footers — those are absorbed into the background.
  3. `layout` is an integer 12-column grid {x,y,w,h}.
     x_grid = round(pixelX / (imageWidth/12)); y_grid = round(pixelY / 33).
     The sum of w on the same row must be <= 12.
     `layout` must enclose the title AND the data beneath it as one region.

  4. [bbox — CRITICAL FOR ACCURACY] For every component, also output a normalized
     bounding box `bbox` measured against the FULL image:
     bbox = {"x": left/imageWidth, "y": top/imageHeight,
             "w": width/imageWidth, "h": height/imageHeight}.
     All values are decimals in 0.0..1.0, with x+w<=1.0 and y+h<=1.0.

     Follow these bbox rules precisely — bbox accuracy is the most important output:
     a. The report content sits on a white card that is INSET from the image edges
        (there is a light gray/blank margin around the card, usually ~6-10% on each
        side). NEVER let any bbox enter that outer margin. Left edges typically start
        around x>=0.07; right edges typically end around x+w<=0.93. The top of the
        first content usually starts around y>=0.08 and the last content ends around
        y+h<=0.97.
     b. `y` (top) must be the ACTUAL top pixel of the component INCLUDING its title.
        Do not place the box above the real content. Measure the top edge carefully —
        a common mistake is starting boxes too high; avoid it.
     c. `h` (height) must cover the FULL vertical extent of the content (title + body).
        Multi-bullet text blocks and tall tables are tall — do not make them too short.
     d. For a two-column layout, the LEFT column occupies roughly x in [0.07, 0.48]
        and the RIGHT column roughly x in [0.50, 0.93]. A left box must satisfy
        x+w<=0.49; a right box must satisfy x>=0.50. Never let a box cross the gutter
        or spill past the card into the gray margin.
     e. Box edges must hug the real content tightly (no slack), but must not clip it.

  5. [SECTION GROUPING — CRITICAL] A heading and the table/chart/list it describes
     form ONE component. Do NOT emit a standalone section_heading for it; put the
     heading text into that component's `title`.
     - "ETF Top 10 보유내역" heading + table  -> ONE holdings_table (title set).
     - "ETF 성과 추이" heading + line chart    -> ONE line_chart (title set).
     - "ETF 리뷰" heading + bullets            -> ONE bullet_list (title set).
     Use section_heading ONLY for a pure divider heading not attached to any data.

  [DATA SPECIFICATION — CRITICAL]
  Each component describes WHAT data it needs via `dataSpec`. NEVER fill in real
  values or column names. `dataSpec` has two parts:

  (A) fields[] : single-value slots (title, base date, department, email — one label = one value)
    { "key": english_key, "label": original_label, "role": "title|meta|scalar|narrative_input",
      "dataType": "string|number|currency|percent|date|text",
      "required": true, "binding": null }
    - Fixed text (section labels, notices) -> role "scalar", dataType "string".
    - Sections whose sentences must be generated by synthesizing several source
      facts (e.g. review / outlook / commentary) -> make a bullet_list and add several
      role="narrative_input" slots. Each slot's `label` states which evidence it needs.

  (B) collection : for tables/charts that draw many rows/columns from one table.
    metric_table, holdings_table, table, line_chart, bar_chart, pie_chart use a collection:
    {
      "enabled": true,
      "table": null,          // which table to read (user maps later; keep null)
      "filterColumn": null,   // search column (e.g. fund code) — user maps
      "filterValue": null,    // search value (e.g. 069500) — user maps
      "columns": [            // each visible result column -> display label + role
        { "label": "종목명", "role": "name",  "dataType": "string", "column": null },
        { "label": "보유비중(%)", "role": "value", "dataType": "percent", "column": null }
      ],
      "orderBy": null, "orderDir": "desc", "limit": 10
    }
    - role values: index | name | value | x | series | label.
    - charts (line/bar): put one role "x" and one or more role "series".
    - pie charts: one role "label" and one role "value".
    - tables: one column per visible header.
    - column/table/filterColumn/filterValue must ALL be null (the user maps them).

  6. Give a lower `confidence` to components you are unsure about.

  [OUTPUT SCHEMA]
  {
   "schemaVersion":"1.0",
   "report":{
     "meta":{"title":str,"pageSize":"A4","orientation":"portrait"},
     "background":{"id":"bg_000","type":"background_image","required":true,"source":"uploaded_sample_image","layout":{"x":0,"y":0,"w":12,"h":50}},
     "grid":{"cols":12,"rowUnit":33,"gap":8},
     "components":[
       {"id":str,"type":<closed set>,"title":str,
        "layout":{"x":int,"y":int,"w":int,"h":int},
        "bbox":{"x":float,"y":float,"w":float,"h":float},
        "dataSpec":{"fields":[...], "collection":{...}|null},
        "confidence":float}
     ]
   }
  }

  Scan the image top->bottom, left->right and reflect the column layout.
  Each bbox (normalized 0..1) must tightly and accurately enclose the real content
  region per rule 4 (correct top, full height, inside the white card, respect the gutter).
  Output ONLY the JSON object described above.

fewshot: |
  [Reference: expected IR excerpt for an ETF issue report — labels/roles only; values/columns are null. bbox values are tightly fitted to the white card, respect the two-column gutter, and have correct tops and full heights.]
  {
   "schemaVersion":"1.0",
   "report":{
    "meta":{"title":"Kodex ETF 이슈 리포트","pageSize":"A4","orientation":"portrait"},
    "background":{"id":"bg_000","type":"background_image","required":true,"source":"uploaded_sample_image","layout":{"x":0,"y":0,"w":12,"h":50}},
    "grid":{"cols":12,"rowUnit":33,"gap":8},
    "components":[
     {"id":"cmp_meta","type":"metadata_text","title":"발행 정보","layout":{"x":1,"y":5,"w":3,"h":3},
       "bbox":{"x":0.11,"y":0.20,"w":0.26,"h":0.07},
       "dataSpec":{"fields":[
         {"key":"base_date","label":"기준일","role":"meta","dataType":"date","required":true,"binding":null},
         {"key":"department","label":"발행 부서","role":"meta","dataType":"string","required":true,"binding":null},
         {"key":"email","label":"연락처 이메일","role":"meta","dataType":"string","required":true,"binding":null}
       ],"collection":null},"confidence":0.9},

     {"id":"cmp_title","type":"title","title":"리포트 제목","layout":{"x":4,"y":6,"w":8,"h":2},
       "bbox":{"x":0.34,"y":0.22,"w":0.55,"h":0.05},
       "dataSpec":{"fields":[{"key":"text","label":"리포트 제목","role":"title","dataType":"string","required":true,"binding":null}],"collection":null},"confidence":0.92},

     {"id":"cmp_perf_chart","type":"line_chart","title":"ETF 성과 추이","layout":{"x":1,"y":11,"w":5,"h":6},
       "bbox":{"x":0.11,"y":0.33,"w":0.37,"h":0.27},
       "dataSpec":{"fields":[],"collection":{
         "enabled":true,"table":null,"filterColumn":null,"filterValue":null,
         "columns":[
           {"label":"기간","role":"x","dataType":"string","column":null},
           {"label":"ETF 수익률","role":"series","dataType":"percent","column":null},
           {"label":"BM 수익률","role":"series","dataType":"percent","column":null}
         ],"orderBy":null,"orderDir":"asc","limit":null}},"confidence":0.85},

     {"id":"cmp_hold_table","type":"holdings_table","title":"ETF Top 10 보유내역","layout":{"x":6,"y":11,"w":6,"h":14},
       "bbox":{"x":0.51,"y":0.33,"w":0.38,"h":0.34},
       "dataSpec":{"fields":[],"collection":{
         "enabled":true,"table":null,"filterColumn":null,"filterValue":null,
         "columns":[
           {"label":"No.","role":"index","dataType":"number","column":null},
           {"label":"종목명","role":"name","dataType":"string","column":null},
           {"label":"보유비중(%)","role":"value","dataType":"percent","column":null}
         ],"orderBy":null,"orderDir":"desc","limit":10}},"confidence":0.88},

     {"id":"cmp_review","type":"bullet_list","title":"ETF 리뷰","layout":{"x":1,"y":28,"w":11,"h":5},
       "bbox":{"x":0.09,"y":0.69,"w":0.82,"h":0.12},
       "dataSpec":{"fields":[
         {"key":"performance_summary","label":"성과 요약 근거","role":"narrative_input","dataType":"text","required":true,"binding":null},
         {"key":"market_context","label":"시장 상황 근거","role":"narrative_input","dataType":"text","required":true,"binding":null},
         {"key":"risk_factors","label":"위험 요인 근거","role":"narrative_input","dataType":"text","required":false,"binding":null}
       ],"collection":null},"confidence":0.8},

     {"id":"cmp_outlook","type":"bullet_list","title":"향후 전망","layout":{"x":1,"y":37,"w":11,"h":5},
       "bbox":{"x":0.09,"y":0.83,"w":0.82,"h":0.12},
       "dataSpec":{"fields":[
         {"key":"outlook_points","label":"전망 포인트 근거","role":"narrative_input","dataType":"text","required":true,"binding":null},
         {"key":"defensive_factors","label":"방어 요인 근거","role":"narrative_input","dataType":"text","required":false,"binding":null}
       ],"collection":null},"confidence":0.8}
    ]
   }
  }

retry: |
  Your previous output failed IR schema validation. Fix the errors below and output ONLY the JSON of the same schema again.
  No markdown, no commentary — JSON only. `type` from the closed set, integer `layout`, follow the dataSpec.fields/collection rules, and include a normalized `bbox` per component.
  [VALIDATION ERRORS]
  {errors}
  [PREVIOUS OUTPUT]
  {previous}