What Data for This Category Looks Like
This category of data typically comes from ERP systems, supplier shipping invoices, logistics tracking information, and equipment manufacturer backends. Commonly processed documents include purchase orders, delivery acceptance forms, serial number lists, and invoice materials. Most fields cover equipment model, serial number, purchase order number, delivery date, and supplier name, with units primarily being units, sets, and pieces. Data updates follow single-project delivery cycles, or sync the latest delivery and inventory information weekly. The structure of individual documents varies widely, ranging from plain-text order forms to tabular serial number lists.
What Constraints Do These Characteristics Impose on the Knowledge Base Retrieval and Recall Link
Decentralized data sources lead to inconsistent field naming, so cross-system field matching logic must be adapted to avoid recall failures caused by inconsistent field names. Update frequency requirements mean the knowledge base must support incremental synchronization to avoid resource consumption from full updates. Diverse document structures require compatibility with both plain-text and tabular content retrieval, ensuring core fields such as serial numbers and models can be accurately recalled. The high error risk in core business links requires prioritizing the accuracy of retrieval and recall, to avoid mismatched equipment or missed key acceptance information.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches precise fields such as equipment model and serial number. A threshold that is too low will introduce irrelevant delivery data, while a threshold that is too high will miss valid recall entries |
limit | Top 8 entries | Single-project delivery data volume is limited. Retaining enough candidates for secondary sorting avoids exceeding the context processing limit |
searchMode | Semantic retrieval | Adapts to cross-system field matching, can recognize model abbreviations or aliases, and is more aligned with business scenarios than pure keyword retrieval |
usingReRank | Enabled | Performs secondary sorting of recalled candidate results by relevance to correct mismatches from semantic retrieval |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands the wording of pre-sales questions, for example, associating "delivery time" with the delivery date field of purchase orders |
Three Common Mistakes
- Retrieval and recall results meet requirements, but the final response only returns citation links without generating corresponding content. The cause is that the context splicing rule is not configured, or the recalled document fields are not correctly mapped to the response template.
- Retrieval accuracy is low after importing Feishu documents, and equipment model fields cannot be matched. The cause is that document structured parsing is not enabled, and tabular content in Feishu documents is not correctly extracted as retrievable independent fields.
- The number of results returned by the retrieval API call is less than the configured
limitvalue. The cause is that knowledge base data update is not completed, or some documents have not been correctly vectorized.
How to Verify Correct Configuration
- Upload a single purchase order document, retrieve "the equipment model of this purchase", and check whether the recalled results include the corresponding model field.
- View the knowledge base update log to confirm that the latest delivery acceptance form has completed vectorization synchronization.
- Enable FastGPT v4.17.0 debug mode to view the original results of retrieval and recall, and confirm that the entries filtered by the
similaritythreshold meet business expectations. - Simulate a pre-sales consultation scenario, input "the delivery serial numbers of this batch of equipment", and verify whether the response associates the corresponding content from the serial number list.
The values provided 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.