What data for this category looks like
Data sources primarily include menu documents exported from store operation systems, scanned offline event posters, online marketing activity configuration sheets, and member benefit explanation documents. Update frequency aligns with operational cycles: menus are updated every 1 to 2 weeks, temporary promotion documents are updated 1 to 3 days before activity launch, and member benefit documents are updated quarterly. Document structure includes both structured fields and unstructured content. Structured fields cover activity ID, applicable store scope, benefit effective period, and redemption rules. Unstructured content includes activity promotion copy and dish descriptions. Most field units are yuan, person-times, and hours. Some documents include store longitude and latitude coordinates.
What constraints these characteristics impose on knowledge base retrieval and recall
High-frequency updates to menu and promotion documents require the knowledge base to support incremental synchronization and scheduled full refresh, to prevent recall of expired content. Many structured fields include geographic and time-based attributes. Targeted retrieval conditions must be set for activity ID, applicable stores, and effective periods, to narrow the recall scope. A single menu document may contain multiple pages of dish information, with lengthy unstructured content. Reasonable segmentation rules must be configured to avoid context truncation. Some documents include store longitude and latitude, which can be combined with geographic retrieval for precise matching. Field formats must be unified, otherwise retrieval matching will fail.
How to configure settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
UPLOAD_FILE_MAX_SIZE | 500 MB | Menu PDFs and activity collection documents for hotel and catering businesses are usually under 500 MB per file. This value covers the upload needs of most business documents |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | A single long menu PDF may contain dozens of pages. 600 seconds covers full parsing time and avoids mid-process interruptions |
Segment Length | 800-1200 characters | Single-segment valid information for catering marketing content (such as activity rules, dish descriptions) has moderate length. This range retains complete semantics and avoids context truncation |
Recall Count | Top 6-10 results | In hotel and catering marketing scenarios, users usually need to match 1-3 activities or dishes. This range covers most needs and avoids redundant recalled content |
Similarity Threshold | 0.75-0.85 | Marketing copy has high semantic similarity. This threshold filters low-match irrelevant content while retaining similar activity plans |
Incremental Sync Trigger Method | Triggered by file modification time | Most catering document updates are temporary adjustments. Triggering by modification time reduces invalid synchronization frequency and adapts to high-frequency update scenarios |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require case-by-case analysis. It is recommended to test on your own samples before finalizing settings.
Three common configuration errors
- When uploading large-sized PDF menus or activity collection documents, the upload progress reaches 90% and then displays an "offset out of range" error. Cause: The
UPLOAD_FILE_MAX_SIZEconfiguration is not adjusted to a value suitable for large documents, or network fluctuations during upload cause interrupted chunked uploads. - When calling an API to trigger a workflow, a prompt indicates that the specified knowledge base was not found, or the workflow cannot read the bound knowledge base ID. Cause: The
knowledge_base_idfield is not correctly included in the API request body, or the global variable is not correctly associated with the knowledge base retrieval node of the workflow. - Retrieval results include expired promotional activity content, and the number of valid recalled content items is insufficient. Cause: Retrieval conditions filtered by activity effective period are not configured, or the similarity threshold is set unreasonably, leading to recall of low-match content.
How to verify correct configuration
- Upload a test large-sized PDF menu, check that the upload progress completes without an "offset out of range" error.
- Call the API with a test knowledge base ID, check that the workflow can normally read document content in the knowledge base.
- Initiate a retrieval request, verify that returned results include currently active marketing content and do not include expired content.
- Adjust the segment length and similarity threshold, initiate multiple retrieval requests, and confirm that the semantic completeness and matching degree of returned results meet business expectations.
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.