| """Whole-page text for the agent's context (parsed approach). |
| |
| The agent (models/minicpm_agent) is a text model, so "the page the mechanic is |
| viewing" is given to it as text: a page's parsed elements rendered in reading |
| order, with Picture/Table elements replaced by their MiniCPM descriptions (the |
| same descriptions the chunker splices inline, but here we keep the WHOLE page, |
| not just the chunks that matched a query). Source is the parsed store's |
| parsed.json — [{"page": n, "elements": [{"class", "bbox", "text", |
| "description"?}]}]. |
| |
| Pure functions, no model/GPU dependency. |
| """ |
|
|
| from __future__ import annotations |
|
|
| from core.chunking import FIGURE_CLASS, SKIP_CLASSES, TABLE_CLASS |
|
|
|
|
| def index_pages(parsed_pages: list[dict]) -> dict[int, list[dict]]: |
| """{page_number: elements} so a turn reads parsed.json once and looks up any |
| page it shows without re-scanning.""" |
| return {pg["page"]: pg.get("elements", []) for pg in parsed_pages} |
|
|
|
|
| def page_to_text(elements: list[dict]) -> str: |
| """A page's elements as one text block in reading order. Figures become |
| "[Figure: <description>]" and tables "[Table: <description>] <cells>", so the |
| agent reads what's on the page (and can name a figure/table to circle) |
| without seeing the image. Headers/footers are dropped.""" |
| lines: list[str] = [] |
| for el in elements: |
| cls = el.get("class", "") |
| if cls in SKIP_CLASSES: |
| continue |
| text = (el.get("text") or "").strip() |
| if cls == FIGURE_CLASS: |
| desc = (el.get("description") or "").strip() |
| if desc: |
| lines.append(f"[Figure: {desc}]") |
| elif cls == TABLE_CLASS: |
| desc = (el.get("description") or "").strip() |
| parts = [] |
| if desc: |
| parts.append(f"[Table: {desc}]") |
| if text: |
| parts.append(text) |
| if parts: |
| lines.append("\n".join(parts)) |
| elif text: |
| lines.append(text) |
| return "\n".join(lines).strip() |
|
|
