Part of the travel terms of Nordlicht Reisen GmbH consists of scanned pages, and some PDFs carry price tables as an image (R-03, R-04). The indexer from 5.1 fetches only the text that the file itself contains. The skillset ss-reise turns that into searchable chunks with vectors: read images, split, embed, pass the source along.
Prior knowledge. Why you cut and how headings serve as boundaries is in "AI track, lesson T4.3", embeddings are in "T4.1"; indexers are built in 5.1, index and vectorizer in 5.2. Foreign text as data (T6.2) also holds for a custom skill.
New in this lesson (changes no R number). The skillset ss-reise has the skills OCR, text merge, splitting, embedding (deployment embed) and a custom skill ziele (web API in a container app ca-skills) that names travel destinations from a fixed list for each chunk.
The requirement
All numbers are assumptions of the task.
| No. | Requirement | Type | Measure | Weight | Source |
|---|---|---|---|---|---|
| 5.3-01 | Text in embedded images is searchable. | F | 100 % of the test scans (5 of them) | Must | R-04 |
| 5.3-02 | Chunks have at most 2,000 characters, 200 characters of overlap. | Q | per chunk | Should | T4.3 |
| 5.3-03 | Every chunk carries vector, title, source and language. | F | 4 fields | Must | 2.2, 5.2 |
| 5.3-04 | Travel destinations per chunk are filterable, without a model. | F | fixed list in the code | Should | R-11 |
| 5.3-05 | Ingestion needs no key. | R | 0 keys | Must | R-07 |
| 5.3-06 | Every skill has a named cost driver. | R | table | Should | R-09 |