What the data for this category looks like
Duty-free marketing content data sources include brand duty-free activity manuals, off-island product announcement documents filed with customs, electronic versions of offline store promotional posters, and member marketing rule documents. Update frequency adjusts based on promotional activities and product restock batches, with no fixed cycle. Document formats cover multi-page PDFs with table of contents, structured Excel product lists, and mixed-image-text Word marketing plans. Fields include applicable activity scope, product category, effective and expiration dates, and quota thresholds. Units are mostly date formats, monetary units, and categorized text.
Constraints on Document Parsing and Chunking From These Characteristics
Wide variation exists between document formats from different sources. Sources range from structured Excel product lists to mixed-image-text marketing plans. This requires the parsing module to support both format recognition and OCR extraction. Merged cells in Excel can break field associations. Structured data alignment and repair must be completed before chunking. Time interval fields in marketing documents require retaining contextual connections. This avoids splitting content across activity cycles. Fixed-format customs filing documents require preserving the original order of fields. Business relationships between fields cannot be disrupted.
How to Set Configurations
| Config Item | Recommended Setting | Rationale |
|---|---|---|
PARSE_OCR_ENABLE | Enabled | Duty-free marketing documents include offline promotional posters and graphic pages of off-island policies. OCR is needed to extract text from images |
Chunk size | 800–1000 characters | Duty-free marketing content mostly includes activity rules and product lists. This length preserves the complete logic of a single activity |
CHUNK_OVERLAP | 100–150 characters | Activity rules have contextual connections across chunks. Overlapping characters ensure contextual integrity during recall |
Recall count | Top 6–8 results | Duty-free marketing content has high keyword concentration. This range covers all valid information for relevant activities |
Similarity threshold | 0.75–0.85 | Filters low-relevance recall results while retaining marketing content that matches user queries closely |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Some large customs-filed documents take longer to parse. This threshold prevents mid-parsing interruptions |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by document format, data volume, and business rules. Specific issues require targeted analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Mistakes
- Symptom: Query content exactly matches a document chunk in the knowledge base, but no corresponding result is recalled. Recalling works normally after creating a new knowledge base. Cause: The
Similarity thresholdis not configured correctly. An overly high threshold filters matching results, or theRecall countvalue is too low to cover matching content. - Symptom: After uploading multiple duty-free marketing documents, parsed results do not distinguish between individual documents. All content is merged into a single chunk. Cause: The parsing configuration for splitting by document is not enabled. This causes content from multiple documents to be mixed during chunking, with no correspondence to original files.
- Symptom: After uploading a complete duty-free marketing PDF document, retrieved question-answering content is scattered and weakly connected. Cause: The
Chunk sizesetting does not match the document content density. This prevents chunks from retaining complete business logic context.
How to Confirm Configurations Are Set Correctly
- Upload a typical duty-free marketing Excel document. Check if parsed fields are complete. Confirm that structured data automatic parsing configuration is active.
- Upload a promotional poster PDF with images. Check if parsed results include OCR text from images. Confirm that OCR configuration is enabled.
- Run a test query. Verify the number and matching degree of recalled results. Adjust corresponding configuration item values to meet business needs.
- Upload multiple different types of duty-free marketing documents. Confirm that chunking results for each document are independent, with no cross-document merging.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.