What the data for this category looks like
Data sources include cross-border financial supporting duty-free product marketing manuals from financial institutions, in-store duty-free event display materials, online financial mall duty-free product detail pages, and member-exclusive duty-free event marketing script libraries. Update schedule: regular categories are updated monthly, promotional periods are updated weekly, and holiday event updates start 1-2 weeks in advance. Document structures typically include product name, duty-free selling price, purchase restriction rules, target audience, activity validity period, redemption method, plus marketing script variants for different scenarios. Fields include sku_id, tax_free_price, valid_period, activity_tag, with units of yuan, piece, day.
Constraints on knowledge base retrieval and recall
Dispersed multi-source materials require support for batch import and parsing of multiple formats, including pdf, docx, markdown and other document types. High-frequency update schedules require incremental indexing functionality to avoid resource waste from full repeated indexing. Mixed document structures of product details and marketing scripts require segment lengths adapted to varying text lengths while preserving semantic integrity. Clear business fields require support for field-level retrieval filtering to ensure recalled content only includes currently valid duty-free marketing information. Colloquial marketing script variants require selection of natural language-adapted embedding models to improve the accuracy of semantic recall.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
UPLOAD_FILE_MAX_SIZE | 100 MB | Duty-free marketing materials often include high-resolution product images and long promotional posts. This value supports batch import requirements for multiple formats |
PARSE_FILE_INCREMENTAL | Enabled | High-frequency content updates during promotional periods. Incremental parsing avoids repeated indexing of full datasets |
chunk_size | 800–1200 characters | Mix of marketing scripts and product details. This segment length preserves semantic integrity and avoids truncating key purchase restriction rules |
similarity_threshold | 0.75–0.85 | Requires accurate recall of marketing scripts and product information. This range balances recall coverage and accuracy |
top_k | Top 8–12 results | Single marketing content has multiple variants. Additional recall covers script needs for different scenarios |
rerank_top_n | Top 3–5 results | Filters redundant results and retains marketing materials most relevant to user queries |
The parameter values provided on this page are general recommendations used as a starting point for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require individual analysis. It is recommended to test on your own samples before finalizing settings.
Three common configuration errors
- Creating a knowledge base prompts "index creation timed out", and the interface freezes during initialization. The cause is failure to adjust the
PARSE_FILE_TIMEOUT_SECONDSparameter, and failure to adapt to long documents for batch import and high-frequency updated marketing materials. - Search results include expired duty-free event information that does not match the valid content set in the knowledge base. The cause is failure to configure retrieval filtering rules based on fields such as
valid_periodandactivity_tag, and failure to limit the timeliness and activity attributes of recalled content. - After replacing a custom embedding model, indexing completes but search tests return "parameter format error". The cause is failure to verify that the output dimension of the custom model matches the dimension required by the platform, causing the recall link to fail to match semantic vectors normally.
How to confirm configuration is properly set
- Upload a single mixed document containing product details and marketing scripts, check if parsed segment fields are fully extracted, and confirm that the
chunk_sizeconfiguration does not truncate key purchase restriction rules and validity period information. - Initiate queries including "duty-free product purchase restriction rules" and "currently valid activities", verify that search results only return compliant content, and confirm that field filtering configuration is active.
- After replacing a custom embedding model, perform batch incremental indexing, check if the indexing progress bar advances normally with no error prompts.
- Adjust the
similarity_thresholdparameter, verify that the number of search results changes as expected with the threshold, and confirm that the similarity configuration is active.
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.