Tencent Cloud Real-Time Document Extraction Agent

What problem it solves

In office workflows, useful information often lives in contracts, invoices, and reports as images or single PDF pages rather than structured records. Manual entry is slow and error-prone, especially when key-value pairs and table fields are mixed together. Tencent Cloud Real-Time Document Extraction Agent addresses this by accepting an image or single PDF page plus the field names you care about, then returning structured JSON.

How it works

It calls the Tencent Cloud OCR real-time document extraction API and targets scenarios with roughly 30-second latency and about 2000 input/output tokens. Key inputs include:
- ImageBase64 or ImageUrl, with support for PNG, JPG, JPEG, BMP, and PDF.
- ItemNames, the list of fields to extract; each field uses KeyName, uses KeyType to distinguish KV pairs (0) and table fields (1), and can include KeyPrompt for field descriptions.
- PdfPageNumber, used to select a single PDF page when the input is a PDF.

The response is not raw OCR text but structured data organized by GroupInfo, LineInfo, and ItemInfo. Field values may include AutoName, AutoContent, coordinates, and page information, which makes it easier for downstream code to consume without regex-based parsing.

Boundaries and notes

This skill is best for real-time field extraction rather than general long-document intelligence. Keep these constraints in mind: encoded image/PDF size should be under 10M, pixel dimensions should be within 20 to 10000px, and PDF input is limited to single-page recognition. ItemNames must not be empty, KeyName must not be empty, KV and table fields have separate limits, and the default request rate is about 20 QPS. If the service is not enabled, download fails, image decoding fails, or the account is in arrears, inspect ErrorCode and ErrorMessage in the response.