What the data for this category looks like
Data for this category comes primarily from cloud billing systems, tag data, order systems, and resource monitoring tools. Daily generated documents include billing details, cost allocation tables, resource utilization reports, and optimization suggestion sheets. Fields cover resource identifiers, billing cycles, actual costs, tag key-value pairs, allocation ratios, and similar items. Common units include yuan, GB, hours, and other standard units. Data update frequency is typically daily or monthly. Some high-frequency resource monitoring scenarios sync latest data hourly.
What constraints these characteristics impose on knowledge base retrieval and recall
The multi-system source of this category’s data creates scattered field dimensions. Retrieval requires precise matching of associated fields such as resource identifiers and tag key-value pairs to avoid mixing unrelated resource cost data into results. Different documents have varying update frequencies. Recall logic must distinguish display priorities between daily resource monitoring data and monthly billing data, and prioritize returning the latest valid data. The long-text structure of detail documents requires careful control over chunking and recall scope to retain key information such as cost allocation and exception annotations. Business requirements for abnormal cost troubleshooting and cost explanation also demand high-confidence recall results, to avoid missing cost entries that may trigger customer disputes.
How to configure settings
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Cloud billing data has strong field correlation. A threshold that is too low will introduce cost data from unrelated resources, while a threshold that is too high may miss matching abnormal billing entries. |
limit | Top 20 entries | Single cloud billing detail entries contain multi-dimensional fields. Too many recall results will exceed the context window, while too few will fail to cover all allocation logic. |
usingReRank | Enabled | The tag and cost association logic of cloud bills is complex. A reranking model can filter low-correlation allocation table entries and improve the accuracy of pre-sales Q&A. |
datasetSearchUsingExtensionQuery | Enabled | Customer inquiries often include vague expressions such as "last month's costs for a certain server". Extended queries can associate tags and resource IDs to match corresponding billing documents. |
searchMode | Hybrid vector retrieval + keyword retrieval | Covers both precise resource ID matching and fuzzy cost description retrieval, adapting to different question scenarios for technical support and pre-sales. |
collectionIds | Configured by business line grouping | Cloud billing data is isolated by business line. Specifying collection IDs prevents recall of unrelated data across business lines. |
Three common misconfigurations
- Symptom: After enabling the
usingReRankand knowledge base index reranking switches in the interface, online recall test results do not show changes in reranking order. Cause: Reranking calls were not enabled simultaneously during the large model invocation phase. Only configuring reranking during the indexing phase will not take effect. - Symptom: After configuring
limitto 3000, the large model does not receive context content during online testing. Cause: The system has an implicit limit on the total number of tokens for a single recall context. The total token count of 3000 detail entries exceeds the window upper limit, causing the context to be automatically cleared. - Symptom: After setting the chunk repetition rate parameter, duplicate billing detail entries still appear in knowledge base retrieval results. Cause: A reasonable overlap length was not configured for long-text chunking of cloud bills, causing the repetition rate calculation logic for adjacent chunks to not match the structure of detail documents.
How to confirm configurations are complete
- Access the FastGPT v4.17.0 knowledge base online test interface, input an inquiry statement containing a specific resource ID, and verify if returned results include the corresponding billing detail document.
- Enable the
usingReRankswitch, then compare the sorting of retrieval results before and after enabling to confirm that high-correlation cost allocation table entries are displayed first. - Adjust the
similaritythreshold, observe the matching accuracy of retrieval results, and confirm that the threshold meets the needs of the current business scenario. - View the retrieval log to confirm that
datasetSearchUsingExtensionQueryhas taken effect, and that extended query terms have been associated with tag and resource ID fields.
The values provided are common starting points and should be measured 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.