What data in this category looks like
Data for this category primarily comes from cloud audit logs, IAM permission systems, CMDB resource inventory libraries, and security baseline scan results. Daily output documents include compliance checklists, permission audit reports, resource inventory lists, and rectification tracking forms. The data updates on a monthly or quarterly cycle. Each individual document contains fields such as resource ID, permission configuration items, compliance status, and rectification records. There is no fixed uniform unit, and updates only occur alongside cloud resource changes and audit cycles.
What constraints do these characteristics impose on the reranking and hybrid retrieval link
Cloud resource compliance auditing data mostly consists of structured or semi-structured compliance documents. Fields focus on permissions, configurations, and compliance status. Update cycles run monthly or quarterly, and individual documents have large file sizes. Hybrid retrieval must balance accurate matching of resource IDs, permission anomaly descriptions, and compliance clauses in user queries. The reranking link must prioritize core keyword association. Because data update frequency is low, retrieval must prioritize the latest audit report versions, and avoid recalling expired rectification records. Additionally, field structures vary slightly across different documents. Reranking must distinguish weights for different document types, to prevent irrelevant resource inventory entries from interfering with compliance-related results.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Reranking model configuration | Qwen/Qwen3-Embedding-8B, configure only once | This model supports embedding and reranking tasks. Identical configurations will overwrite existing settings, so uniqueness must be ensured |
| Single search recall limit | Top 8–12 entries | The number of compliance audit documents is relatively manageable. Excessive recall increases reranking computation load |
| Reranking result similarity threshold | 0.72–0.80 | Low-correlation resource inventory entries must be filtered out, to retain results strongly related to compliance and permissions |
| Document type weight setting | Permission audit report > Compliance checklist > Rectification tracking form > Resource inventory | Most user queries focus on permission anomalies and compliance clauses, so retrieval weights for core documents must be increased |
| Retrieval timeout duration | 500–700 seconds | Compliance documents have large file sizes, so sufficient time must be reserved for retrieval and reranking processing |
| Hybrid retrieval toggle | Enabled | Semantic matching and keyword matching must be combined, to improve recall accuracy for compliance-related content |
Three common mistakes
- Phenomenon: The interface prompts that the reranking model cannot run, or model call failure records appear in logs. Cause: The reranking model was added without following the unique configuration rule, or a model that only supports embedding tasks was used.
- Phenomenon: The reranking link returns false during the question-answering flow, with no valid reranking result entries. Cause: No available reranking model was configured, or the similarity of all recalled results is lower than the set filter threshold.
- Phenomenon: Reranking results do not hit the specific cloud resource ID mentioned in the user’s query. Cause: The reranking link uses rewritten query text, and does not retain resource identification keywords from the original query.
How to confirm the configuration is complete
- Enter the model configuration page of FastGPT v4.17.0, check that the reranking model is configured only once, and the name matches the deployed model.
- Initiate a simulated compliance audit-related query, verify that the number of recalled results in the retrieval link meets the preset limit.
- View the running logs of the reranking link, confirm that the model call is successful and returns a valid sorted result list.
- Compare the original recalled results with the reranked results, confirm that highly relevant compliance documents are prioritized.
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.