What the data for this category looks like
Typical data for this category comes from order design systems, sheet material and wood batch records, CNC cutting and drilling equipment logs, ERP/MES systems, and on-site inspection records. Daily documents include order breakdown drawings, cutting lists, edge banding and drilling programs, coating process specifications, hardware matching lists, trial assembly and packaging records. Data updates trigger when custom orders are created, daily production data is synchronized, and each batch of materials arrives. Most documents include fields such as panel size, production batch number, and process steps. There is no unified fixed format, and adjustments are made based on individual order requirements.
What constraints these characteristics impose on knowledge base retrieval and recall
Custom order-driven production creates large volumes of highly personalized documents per batch. The retrieval link must support multi-document associated recall to match process requirements for corresponding panels. Data updates from multiple sources have varying frequencies. An incremental update mechanism sorted by order or batch must be supported. Documents lack a unified format, with significant differences in paragraph lengths. The retrieval link must adapt to chunking and recognition of irregular text. Accurate matching of fields like panel size and batch number directly impacts production outcomes. Retrieved results must be strictly linked to corresponding production parameters.
Configuration Settings
This configuration is adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.88 | Adapts to the personalized characteristics of furniture and wood product process documents, balances recall volume and result accuracy, and prevents irrelevant panel parameters from being included |
limit | Top 6–10 results | A single batch of custom orders involves many panels and process documents. A single recall covers core production and exception handling needs adequately |
usingReRank | Enabled | Process document paragraphs have large differences. The re-ranking step filters low-relevance results and improves content accuracy when drafting exception reports |
datasetSearchUsingExtensionQuery | Enabled | Adds retrieval extensions for exclusive fields such as order numbers and panel batches to match production data from multiple sources |
collectionIds | Bind corresponding collections by production document category | Differentiates document types such as order breakdown drawings, process specifications, and hardware lists to avoid interference from irrelevant cross-category content |
Three Common Errors
- Phenomenon: After upgrading a private deployment version, clicking the knowledge base file details displays "Invalid dataset file key". Reason: The key cache associated with the dataset file was not updated synchronously during the upgrade, causing the system to fail to recognize the legal identifier of the original file.
- Phenomenon: After knowledge base chunking, retrieval results show mismatched panel sizes and process steps. Reason: The model paragraph recognition function was not enabled, or when the maximum paragraph depth is set to 3 and the maximum chunk size is set to 1000, the hierarchical structure of process documents was not adapted, causing chunking to break the association between processes and corresponding panel parameters.
- Phenomenon: Unable to synchronize real-time data from Feishu multidimensional tables when pulling external document libraries via API. Reason: The API call permissions of the external data source were not configured correctly, or the dataset extended retrieval function was not enabled, making it impossible to associate hardware lists or cutting data from multidimensional tables.
How to Verify Proper Configuration
- Enter the panel size of a custom order during retrieval testing, and confirm whether the recalled results include the corresponding cutting list and coating process documents.
- View the knowledge base chunking log to confirm that chunking was completed using model paragraph recognition, and that the maximum paragraph depth and chunk size match the preset configuration.
- Call the retrieval API to verify that the returned results include production batch data from the ERP system and supporting information from the external document library.
- Check the system log to confirm that no abnormal errors related to dataset file keys have occurred.
The values provided are common starting points and should be measured against the reader's own samples.