fix: improve page index descriptions (#17268)

api/db/init_data/compilation_templates/page_index.yaml

@@ -4,7 +4,25 @@ config:
   kind: page_index
   entity:
     description: >-
-      You are a robust Table-of-Contents (TOC) extractor.
+      You are a robust table-of-contents extractor with source-grounded
+      compressed descriptions. Extract each heading as an entity. The
+      entity's `name` must be the exact clean heading text. For each heading,
+      write a strongly compressed summary of only the content belonging to
+      that heading: normally 1–2 sentences, aiming for about 50–100 English
+      words or 80–180 Chinese characters when the source is long enough. Keep
+      only the 2–4 most important facts, prioritizing key
+      numbers, dates, names, locations, and conclusions. Do not copy long
+      passages verbatim and do not reduce the description to a generic label.
+      For detailed lists, give the overall result and only the most important
+      examples. Do not write meta descriptions such as `heading for ...`,
+      `section heading`, or `subheading for ...`. Do not include image
+      captions, `Main article` links, or unrelated tables and lists unless
+      they are the actual content of the heading. If a parent heading has no
+      direct body before the next child heading, generate a short higher-level
+      overview from the available child-section content; do not copy the child
+      description verbatim. This parent overview is needed for later
+      cross-chunk merging. Only use an empty description when there is no
+      usable source content at all. Do not invent content.
     fields:
       - type: title
         description: the heading text (clean, no page numbers or leader dots)
@@ -15,8 +33,10 @@ config:
           - "title" must be non-empty (or exactly "-1").
           - If any part of a chunk has no valid heading, output that part as {"title":"-1", ...}.
           - If a chunk contains multiple headings, expand them in order:
-              - Each heading → {"title":"...","chunk_id":"<chunk_ID>"}.
-          - When ambiguous, prefer "-1" unless the text strongly looks like a heading.
+              - Each heading → {"title":"...","chunk_id":"<chunk_ID>","description":"a compact summary of only that heading's section"}.
+          - A parent heading with no direct body must still receive a short higher-level overview based on its child-section content for later merging; do not copy a child's summary verbatim.
+          - Prefer 1–2 concise sentences, aiming for about 50–100 English words or 80–180 Chinese characters when the source is long enough; preserve only the most important facts, dates, numbers, names, and conclusions.
+          - When ambiguous, prefer `-1` unless the text strongly looks like a heading.
           - Keep language of "title" the same as the input.
           - Prefix like following must be titles: 第x章, 第N条, 第N节, 1, 1.1, 1.1.1 ...
   relation: