What this category of data looks like
Data for this category comes from ad delivery platforms, landing page assets, product qualification documents, and advertising law compliance clauses. Commonly processed documents include ad creative copy, landing page screenshots, audit opinion sheets, and modification record ledgers. Single Word documents can reach 100,000 Chinese characters, and Excel datasets can exceed 15,000 rows. Data update frequency is daily or real-time. Core fields include creative copy content, landing page URL, audit status, and compliance judgment result. No custom units are used. This scenario uses FastGPT v4.17.0.
What constraints do these characteristics impose on the "document parsing and chunking" link
Single large-volume documents increase parsing duration. Word documents over 100,000 Chinese characters and Excel datasets with 15,000 rows can cause vectorization errors or parsing timeouts without properly configured chunking rules. Real-time updated creative data requires parsing processes with low latency, to avoid ad rejection caused by delayed compliance audits. Different document types (text, tables, screenshots) need tailored parsing logic. Landing page screenshots require additional OCR processing. Mixed-format asset packs raise parsing failure rates. Chunk length must match the context limits of vector models, to avoid single chunks exceeding processing ranges and reducing the accuracy of violation point annotation.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
UPLOAD_FILE_MAX_SIZE | 10000 MB | Meets upload requirements for single Word documents with 100,000 Chinese characters and Excel datasets with 15,000 rows |
PARSE_FILE_TIMEOUT_SECONDS | 600–900 seconds | Covers full parsing time for large documents, prevents mid-process timeout interruptions |
chunkSize | 800–1200 characters | Matches context window limits of mainstream vector models, while preserving semantic integrity of creative copy and audit opinions |
chunkSplitter | Split by paragraph + fixed character count | Adapts to natural paragraph structure of ad creative copy and audit opinions, prevents violation annotations from being broken by chunking |
customPdfParse | Enable OCR adaptation mode | Processes PDF documents containing landing page screenshots, extracts creative copy and landing page URL content from images |
chunkSettingMode | Manual configuration | Independently adjusts chunking parameters for large documents, adapts to long text and tabular data specific to this category |
Three common errors
- Phenomenon: Vectorization errors occur after generating 1000+ chunks when uploading files larger than 10M, and recovery fails after repeated retries. Cause:
chunkSizeand chunking parameters are not adjusted for large documents, causing single chunk content to exceed vector model processing limits. Chunk verification mechanism is not activated. - Phenomenon: Fields are empty or formatted chaotically after parsing 15,000-row Excel data. Cause: ChunkSplitter rules adapted to tabular data are not configured. Default chunking logic cannot recognize multi-row and multi-column structures, leading to loss of key compliance information during data splitting.
- Phenomenon: No valid text content is returned after parsing landing page screenshot documents. Cause: OCR adaptation mode for
customPdfParseis not enabled. Only PDF metadata is extracted instead of recognizing creative copy in images, making landing page consistency verification impossible.
How to confirm the configuration is correct
- Upload a single Word document with 100,000 Chinese characters, check if parsing completes within the time set by
PARSE_FILE_TIMEOUT_SECONDS. - Extract one ad creative copy, manually split it to the character length set by
chunkSize, verify that semantic integrity is not compromised. - Upload a landing page screenshot file, confirm that the parsing result includes the text content in the image, verifying that the
customPdfParseconfiguration takes effect. - Import 15,000-row Excel data, check the chunk list, confirm that data is split into reasonable chunks matching the
chunkSizesetting, with no fields lost.
The values listed above are common starting points and should be evaluated against the user's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.