What data looks like for this category
Daily documents for this category fall into four types: incentive rules, settlement details, appeal records, and payment lists. Data sources include content data, advertising revenue sharing systems, financial systems, and account backends. Update frequency is monthly or aligned with single event cycles. Document fields include creator identifier, revenue amount, settlement cycle, appeal status, payment progress, and more. Units involve yuan, number of occurrences, cycle number, and others. Data formats vary slightly across different sources, so fast retrieval is needed to integrate and locate abnormal information.
What constraints do these characteristics impose on reranking and hybrid retrieval
Data for this category is scattered across documents from multiple system sources, and update rhythms fluctuate with event cycles. Information weights vary significantly across different documents: incentive rules are general explanatory documents, while settlement details and appeal records are core business data. Hybrid retrieval must cover both rule content matched via keywords and detail data matched via vectors, to avoid missing key business information. The reranking step must prioritize sorting core settlement-related fields, avoid interference from irrelevant rule documents during operations troubleshooting, and adapt to data update frequencies to ensure retrieval results reflect the latest settlement and appeal status, reducing information screening costs for on-call operations staff.
How to set configurations
| Config Item | Suggested Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Adapts to field matching accuracy for settlement details, filters out irrelevant general rule documents and retains core business data |
limit | Top 10 results | Controls the number of returned results, adapts to the need for fast anomaly location during on-call operations, and avoids interference from redundant information |
searchMode | hybrid | Hybrid retrieval mode, integrates vector retrieval and keyword retrieval, and covers both rule-type and detail-type documents |
usingReRank | Enabled | Reranks recalled results by settlement key field priority, prioritizing display of core business content such as appeal records and settlement details |
Three common misconfigurations
- Empty returned results or missing reference fragments after enabling reranking. The cause is setting the
similaritythreshold too high, which filters out weakly relevant but critical detail data that fits the settlement scenario. - Incentive rule documents make up an overly high proportion of retrieval results, interfering with anomaly troubleshooting. The cause is not setting
searchModeto hybrid; using a single retrieval mode fails to cover keyword-based rule content. - The number of retrieved results exceeds expectations, increasing ticket processing time. The cause is not configuring a reasonable
limitparameter, which leads to recall of excessive redundant historical settlement data.
How to confirm correct configuration
- Check the FastGPT v4.17.0 service startup logs to confirm that the AIPROXYAPIENDPOINT (root address of the AI Proxy service, for example http://aiproxy:3000) and AIPROXYAPITOKEN (AI Proxy administrator token, consistent with the AI Proxy server configuration and not empty) environment variables are configured correctly, with no startup verification failure prompts.
- Manually enter settlement-related query terms, such as "2024 creator appeal records", and check if returned results first include document fragments related to settlement details and appeals.
- Compare retrieval results before and after enabling reranking, and confirm that results matching settlement key fields are sorted higher after reranking.
- Verify that the number of retrieval results matches the configured
limitparameter, with no excessive redundant content.
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.