What data for this category looks like
Data for this category comes from cloud billing systems, tag data, order systems, and resource monitoring. Daily documents include billing details, cost allocation sheets, resource utilization reports, and optimization suggestions. Billing details contain billing items, billing cycles, and cost amounts. Cost allocation sheets contain resource attribution tags, allocation ratios, and post-allocation costs. Resource utilization reports contain resource IDs, CPU utilization, and memory utilization. Optimization suggestions contain abnormal cost items and optimization directions. Data update frequency is daily or monthly. Fields include resource identifiers, cost values, utilization metrics, and no additional statistical percentages.
What constraints do these characteristics impose on reranking and hybrid retrieval
Multi-source structured data for this category requires hybrid retrieval to cover documents from different data sources such as cloud billing, tags, and orders. It must match the field association logic of different documents. Data update frequencies vary between daily and monthly, so the recall range must be adjusted to adapt to documents with different update rhythms. Fields include business-related fields such as costs, utilization rates, and tags. Reranking must prioritize core ticket scenarios such as matching abnormal costs and tag attribution verification. The accuracy of retrieval results directly affects ticket response efficiency. Low-correlation historical billing data must be filtered to avoid customer billing disputes. Additionally, the operations duty scenario requires controllable retrieval response times to avoid excessive waiting that delays duty processing progress.
This page applies to FastGPT v4.17.0
How to set configurations
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
| Reranking Toggle | Enabled | Documents for cloud billing and cost accounting contain multi-field structured content, requiring filtering of low-correlation billing entries |
| Similarity Threshold | 0.72–0.88 | Filter billing details and cost allocation sheet data that meet the matching standard for ticket questions, avoiding irrelevant billing items |
| Single Recall Limit | Top 6–10 entries | Cover core billing details, allocation sheets, and resource utilization report data, controlling retrieval response time |
| Retrieval Mode | Hybrid Retrieval | Adapt to hybrid retrieval requirements for multi-source structured documents, matching field characteristics of different data sources |
| Reranking Recall Count | Top 3–5 entries | Prioritize displaying data strongly related to ticket questions, such as abnormal costs and tag verification |
Three common misconfigurations
- Phenomenon: Hybrid retrieval requests take longer than 15 seconds to return, and the interface shows a timeout prompt. Cause: The value range of the single recall limit is not restricted, and excessive historical billing data is recalled, increasing the computational load of retrieval and reranking.
- Phenomenon: Only 1 result is returned after reranking, which does not match the ticket question. Cause: The similarity threshold is incorrectly set too high, filtering eligible allocation sheet and utilization report data.
- Phenomenon: Retrieval results do not include cost allocation data related to resource tags. Cause: The reranking toggle is not enabled, or the hybrid retrieval mode is not configured, only recalling billing data from a single data source.
How to confirm configurations are correct
- Check that the
AIPROXY_API_ENDPOINT(root address of the AI Proxy service) andAIPROXY_API_TOKEN(AI Proxy administrator token) of the App/Admin server are correctly configured, ensuring no verification failure prompts appear during service startup. - Enter the retrieval configuration page of the corresponding application, confirming that the reranking toggle is enabled and the retrieval mode is set to hybrid retrieval.
- Initiate a test ticket query, verifying that retrieval results include relevant entries for billing details and cost allocation sheets, and that response times meet expectations.
- Adjust the similarity threshold, verifying the correlation filtering effect of retrieval results, confirming that low-correlation data is effectively filtered.
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.