What the data for this category looks like
This category's data mainly comes from advertising platform backends, customer qualification repositories, contract systems, and publicly available regulatory rules. The core documents processed daily include four types: advertising review forms, qualification attachments, delivery records, and rectification notices. Data updates are triggered upon completion of a single advertising campaign or on a monthly cycle, with no fixed high-frequency update rhythm. Most documents are a combination of structured forms and attachments, containing fields such as delivery subject, compliance review status, qualification validity period, and rectification deadline. There is no unified standard format, and some attachments are scanned documents or unstructured text.
What constraints these characteristics impose on reranking and hybrid retrieval
Dispersed data sources and inconsistent formats require hybrid retrieval to support both structured fields and unstructured attachments. Clear association matching rules for multi-source data must be defined. Update frequency is tied to campaign triggers. Retrieval must prioritize matching the latest campaign-related documents to avoid calling expired data. Documents include time-sensitive fields such as qualification validity period and rectification deadline. Reranking must prioritize content with strong timeliness. There is a strong association between delivery records and rectification notices. Hybrid retrieval must support cross-document associated recall to ensure a complete compliance evidence chain. Compliance audit trail requirements demand that retrieval results are traceable. Reranked results must retain original data source markers to meet subsequent audit needs.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled | Compliance audit trail scenarios require prioritizing recall of the most relevant compliance evidence. Reranking filters low-correlation redundant results and improves retrieval accuracy |
limit | Top 8-12 entries | The volume of compliance documents grows with each advertising campaign. Limiting the number of returned entries avoids context overflow, while covering the volume of documents required for a core evidence chain |
similarity | 0.75-0.85 | This range balances recalling compliance-related content and avoiding missing marginal evidence. It filters irrelevant general advertising content and retains results directly related to compliance review |
searchMode | Hybrid Retrieval | This category's data includes structured forms and unstructured attachments. Hybrid retrieval matches both field keywords and text semantics, fully covering different forms of compliance evidence |
Three Common Misconfigurations
- Symptom: Retrieval returns no results after enabling the reranking model within an application, but normal results are returned during internal knowledge base testing. Cause: The reranking model is not associated with the knowledge base retrieval link of the corresponding application. Reranking parameters are only configured in the knowledge base test environment.
- Symptom: The number of retrieval results does not match the configured
limitparameter, and a large number of redundant non-compliance documents appear. Cause:usingReRankis not enabled, or thesimilarityparameter is set too low, resulting in recall of low-correlation content. - Symptom: Offline reranking model configuration fails after upgrading FastGPT to v4.17.0, with an interface prompt indicating configuration abnormality. Cause: The dependency package version of the reranking model does not match the FastGPT core package version during offline deployment, or the local storage path of the model is not correctly configured.
How to Confirm Proper Configuration
- The retrieval debugging page of the FastGPT application is accessed, compliance-related query terms are entered, and whether returned results include core document types such as advertising review forms and delivery records is verified.
- After
usingReRankis enabled, the number of retrieval results and their relevance before and after enabling are compared to confirm that the reranking model has taken effect. - Source markers of retrieval results are checked to confirm that they can be traced back to corresponding data sources such as advertising platform backends and customer qualification repositories, meeting the traceability requirements of compliance audit trails.
- The
similarityparameter range is adjusted, the relevance of retrieval results is verified to change appropriately with the parameter, and configured parameters are confirmed to be suitable for the current business data.
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.