Tencent Cloud Question Marking Agent

Problem

K-12 exam marking is not just handwriting recognition. Real inputs may be full exam images, PDF pages, or single-question screenshots. A production pipeline needs question splitting, item-level judgment, answer comparison, error analysis, and handwriting position output. Traditional setups often separate OCR, question ingestion, and rule-based checks, causing inconsistent fields and difficult coordinate alignment. Tencent Cloud Question Marking Agent wraps this as an async job for both handwritten grading and non-handwritten analysis.

How It Works

It uses Tencent Cloud OCR's Submit and Describe flow: create a job, get JobId, then poll until DONE or FAIL. Input can be ImageBase64 or ImageUrl, supporting common image formats and PDF. Key behaviors:
- Full-sheet marking: splits questions and returns MarkInfos.
- Single-question marking: set BoolSingleQuestion=true to skip splitting; pass ReferenceAnswer when needed.
- Structured results: AnswerInfo returns handwriting, correctness, and analysis, with optional correct answer, knowledge points, and coordinates.
- Optional deep thinking: EnableDeepThink adds deeper reasoning but slows response.

Use it in grading pipelines, homework dashboards, or explanation pages. Note the default limit is 10 questions/minute, PDF is single-page, JobId relies on async status, and MarkInfo may be nested recursively.