This document is based on FastGPT v4.17.0 configuration guidelines. The following content addresses the operations on-duty and internal ticket Q&A scenario for ad account opening audit, and covers specific configurations, constraints, and issue resolution for reranking and hybrid retrieval.
What the Data for This Category Looks Like
This category’s data originates from ad platform backends, the National Enterprise Credit Information Publicity System, customer-submitted account opening application materials, and industry access rules. Daily processed documents include account opening application forms, scanned business licenses, corresponding industry qualification documents, and historical account opening audit records. Data is updated via daily batch synchronization or real-time submission with each individual order. Each data entry includes core fields such as subject name, unified social credit code, applied industry category, and audit status. Document formats vary across different sources.
Constraints Imposed by These Characteristics on Reranking and Hybrid Retrieval
Cross-data-source content requires hybrid retrieval coverage, so multi-source recall channels must be configured. Field differences across document types require the reranking model to support mixed matching of structured qualification fields and unstructured application text. The update rhythm of real-time per-order or daily batch updates requires retrieval indexes to be synchronized regularly, to prevent calling expired audit rules or qualification data. Core fields of individual data entries are concentrated on subject qualifications and industry categories. During reranking, these high-weight fields must be prioritized for association, to improve the accuracy of audit-related retrieval.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Enable Reranking | Enabled | This category requires precise matching of qualifications and audit rules. Reranking filters low-correlation recall results |
| Chunk Similarity Threshold | 0.72–0.85 | Adapts to mixed matching accuracy of structured qualification fields and unstructured application text |
| Single Recall Limit | Top 6–10 entries | Balances retrieval efficiency and context length, adapts to the fast query scenario of operations on-duty |
| Retrieval Mode | Hybrid Retrieval | Covers multi-source content such as ad platform backends and customer-submitted materials, enabling cross-source hybrid retrieval |
| Reranking Recall Count | Top 4–6 entries | Focuses on high-correlation results, adapts to the fast positioning needs of internal ticket Q&A |
Three Common Misconfigurations
- Symptom: No reranked results are displayed during online recall testing, and retrieval results remain in the initial recall order. Cause: Reranking functionality is not enabled in the retrieval configuration, or the reranking model is not associated during the knowledge base indexing phase.
- Symptom: An Invalid URL (POST /v1/rerank) error is returned when calling the reranking model. Cause: The API interface address of the reranking model is configured incorrectly, or the deployed reranking service does not expose the corresponding POST request path.
- Symptom: Non-qualification documents with low correlation account for an overly high proportion in retrieval results. Cause: The similarity threshold is set unreasonably, failing to filter irrelevant content, or the hybrid retrieval mode is not configured to cover core data sources.
How to Confirm Proper Configuration
- Access the FastGPT retrieval test page, input a query related to account opening applications, and verify if returned results have their order adjusted through reranking.
- Review the current retrieval configuration items, confirm the enable reranking switch is turned on, and confirm a compatible reranking model is bound.
- Switch the retrieval mode to hybrid retrieval, and validate that multi-source documents from ad platform backends and customer-submitted materials can be recalled.
- Review retrieval logs, confirm the reranking model call request has been successfully sent and returned valid results.
The values provided are common starting points, and should be evaluated against the samples relevant to the specific deployment scenario.
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.