What the Data for This Category Looks Like
Business handling process data for refined oil sales is sourced from gas station standardized operation manuals, POS system process documents, customer service work order records, and compliance documents from superior units. Two update cycles apply to this data: compliance documents are updated quarterly, and daily operation process documents are updated monthly. Documents are split by business nodes, covering modules including fuel card application, invoice request, abnormal transaction processing, and more. Fields include operation steps, preconditions, required documents, processing time limits (unit: hours), system operation paths, and more. Some fields contain precise descriptions combining numbers and units.
Constraints Imposed on Knowledge Base Retrieval and Recall
A large number of precise descriptions combining numbers and units appear in the documents, such as "processing time limit 2 hours" and "Step 3 insert IC card". Retrieval must precisely match combinations of numbers, text and units to avoid matching deviations caused by token splitting errors. Documents are stored classified by business nodes, so retrieval must recall content targeted to specific nodes to avoid mixing irrelevant results across categories or nodes. Data updates follow a fixed schedule with clear incremental nodes, so the knowledge base must perform incremental updates by business module to avoid resource waste from full synchronization. Fields include short text such as system operation paths, so retrieval must retain context fragments to avoid truncating key operation information.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.82 | Matches precise step descriptions and unit information in refined oil business, reduces the probability of irrelevant recall, and adapts to the embedding model accuracy of FastGPT v4.17.0 |
limit | Top 6 entries | Covers all relevant document fragments for a single business handling node, ensuring complete display of the handling process |
searchMode | Recall by business node classification | Matches the document structure divided by nodes such as fuel card, invoicing, and exception handling, and recalls corresponding content in a targeted manner |
usingReRank | Enabled | Corrects token splitting errors between numbers and text, and solves the problem of extra spaces appearing in content entered without spaces |
datasetSearchUsingExtensionQuery | Enabled | Expands user colloquial inquiries (such as "how to reissue a fuel card") into standard process queries, improving matching accuracy |
collectionIds | Bind exclusive knowledge base collection for refined oil business | Isolates documents from other categories, ensuring that recall results only come from target business data |
Three Common Mistakes
- Phenomenon: Returned document fragments show extra spaces between numbers and text that were entered without original spaces. For example, "Step 3" displays as "Step 3". Cause: The
usingReRankreranking function is not enabled, so token splitting errors during embedding are not corrected. - Phenomenon: Retrieval recalls documents unrelated to refined oil sales, such as administrative office content. Cause: The
collectionIdsconfiguration to bind the exclusive knowledge base collection is not configured, so document data from other categories is mixed in. - Phenomenon: Single retrieval returns too few results to cover complete handling process steps. Cause: The
limitvalue is set too low, so all relevant document fragments for a single business node are not covered.
How to Verify Correct Configuration
- A test document containing a combination of numbers and text, such as "Step 2 insert fuel card", may be uploaded. A retrieval is initiated, and extra spaces in returned results are checked.
- The knowledge base collection binding configuration is reviewed. Only documents related to refined oil business are confirmed to be included, with no content from other categories.
- Different colloquial inquiries, such as "how to reissue a fuel card" and "how to handle abnormal refueling", are initiated. Recalled results are verified to match corresponding business nodes.
- The
similaritythreshold is adjusted, and the matching accuracy of recalled results against business requirements is verified.
The values provided here are common starting points and should be evaluated against local deployment samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.