What This Category of Data Looks Like
This category typically supports specific integrated project delivery. Daily documents include purchase orders, delivery acceptance forms, serial number lists, invoice materials, and more. Data sources cover ERP systems, supplier shipping documents, logistics information platforms, and equipment manufacturer backends. Data updates follow per-order or weekly cycles. Individual documents mostly contain structured fields such as purchase order number, equipment model, serial number, supplier name, delivery date, and acceptance result. Documents like serial number lists are often multi-column structured tables. Column association must be retained to enable accurate retrieval.
Constraints for Knowledge Base Retrieval and Recall
This category differs from general enterprise IT procurement. Multi-column structured documents require retrieval systems to retain field associations, avoiding splitting that breaks column data. Multi-source data sources need to be bound to different datasets to narrow retrieval scope and improve efficiency. The high-frequency update feature requires recall results to sync with the latest delivery information, avoiding use of outdated data. The number of devices involved in a single batch of delivery is limited, but matching accuracy for key fields such as serial numbers is high, requiring a balance between recall scope and precision. In addition, long-text equipment parameter descriptions may exceed the processing limit of conventional retrieval models, so targeted reranking rules need to be configured.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category requires precise matching of key fields such as equipment model and serial number. This range filters out irrelevant results with low similarity |
limit | Top 8 entries | The number of devices involved in a single batch of delivery is usually limited. Excessive results increase the verification burden for warehouse administrators or procurement staff |
searchMode | Vector retrieval + keyword hybrid | Daily documents contain both structured fields (such as serial numbers) and unstructured descriptions (such as acceptance requirements). The hybrid mode balances precise matching and recall scope |
usingReRank | Enabled | The retrieval content of this category often includes long-text equipment parameters. Enabling this optimizes the sorting effect of long content, solving the problem that reranking is not triggered when content is too long |
datasetSearchUsingExtensionQuery | Enabled | Data of this category comes from multiple source systems. Enabling this optimizes the problem expansion ability of cross-source retrieval, improving the effect of associative retrieval between different documents |
collectionIds | Bind corresponding datasets by business module | Data is stored classified by purchase order, delivery acceptance form, and other categories. Binding corresponding datasets narrows the retrieval scope and improves retrieval efficiency |
This configuration is based on FastGPT v4.17.0.
Three Common Misconfigurations
- Symptom: Retrieval return results are sorted chaotically, with relevant content from long documents appearing late. Cause: The
usingReRankconfiguration is not enabled, or long text exceeds the input limit of the reranking model, causing reranking to not trigger. - Symptom: After uploading a multi-column Excel document, only the first two columns are recognized as knowledge base content. Cause: No chunking rule is configured for structured multi-column documents. The default chunking logic only recognizes single-column text and does not retain column association relationships.
- Symptom: Retrieval results cannot accurately extract the serial number information of a specified device. Cause: The
similaritythreshold is set too low, resulting in too many interference results from non-key fields, or thedatasetSearchUsingExtensionQueryconfiguration is not enabled, making it impossible to associate serial number information across documents.
How to Verify Configurations Are Correct
- Upload a typical multi-column Excel document of this category, check if the knowledge base chunking results include all column contents, and confirm that the chunking rule is compatible with structured tables.
- Submit a query containing equipment model and serial number, check if the retrieval results are sorted by relevance, and verify that the
usingReRankfunction is active. - Bind the dataset of a specified business module, submit a query for the corresponding business, confirm that only relevant content within the dataset is returned, and verify that the
collectionIdsconfiguration is active. - Input a long equipment parameter description, check if the retrieval results are sorted by relevance, and confirm that the reranking model is called normally for long content.
The values listed are common starting points and should be evaluated against the reader'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.