What the Data for This Category Looks Like
The core business data for this category comes from advertising platforms, customer relationship management systems, attribution SDKs, and user form submission systems. Daily maintained documents include standardized attribution rule documents, daily updated lead distribution tables, channel effect reports, and abnormal lead checklists. Data fields include channel identifiers, lead unique IDs, conversion timestamps, assigned sales accounts, budget consumption amounts, and more. Update frequencies are mostly daily or weekly.
What Constraints These Characteristics Impose on Reranking and Hybrid Retrieval
Multi-source data access requires hybrid retrieval to support field integration across advertising platforms, CRM systems, and attribution SDKs. Recall logic adapted to multiple data sources must be configured. Document structures include text rules, tabular reports, and structured forms. The reranking link must support input content in different formats. Data update frequency is daily or weekly. Retrieval indexes must sync with the update rhythm to avoid returning outdated attribution rules or lead data. Fields are closely linked. Reranking weights must be adjusted based on the business priority of fields such as channel identifiers and conversion times, to avoid irrelevant leads being included in the result set.
How to Set the Configuration
This configuration is based on FastGPT v4.17.0. Specific configuration items and their rationales are listed below:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Rerank Switch | Enabled | Lead data for this category is closely linked, and reranking is required to filter low-relevance attribution results |
| Similarity Threshold | 0.72–0.85 | Matches the business relevance requirements for core fields such as channel identifiers and conversion times |
| Single Retrieval Limit | Top 10 entries | Controls the number of leads returned per retrieval, adapting to daily troubleshooting and analysis scenarios |
| Recall Mode | Cross-source hybrid recall | Adapts to the retrieval needs of multiple data sources (advertising platforms, CRM, attribution SDKs) for this category |
| Rerank Trigger Length Threshold | 800–1200 characters | Matches the average content length of daily documents for this category (attribution rules, channel reports) |
| Index Sync Cycle | Daily/Weekly | Matches the update frequency of data for this category, avoiding retrieval of outdated content |
Three Common Mistakes
- Symptom: After the rerank model is deployed, it can pass independent tests, but the rerank mark in retrieval results is false, and the returned results are not sorted by reranking. Cause: The call configuration for the rerank model is not associated with the current retrieval task, or the rerank switch is not enabled correctly.
- Symptom: When the length of knowledge base retrieval content exceeds the threshold, the rerank model is not automatically called. Cause: The rerank trigger length threshold is set too low, and the threshold parameter is not adjusted for long-text attribution rules and channel reports in this category.
- Symptom: The rerank model cannot function normally in large model dialogue scenarios, but runs normally in separate knowledge base tests. Cause: The rerank switch is not enabled in the retrieval configuration for large model dialogues, or the retrieval scope does not cover the core data sources of this category.
How to Confirm the Configuration Is Complete
- Navigate to the retrieval configuration page in FastGPT, confirm that the rerank switch is enabled, and the rerank trigger length threshold matches the average content length of documents for this category.
- Submit a test query that includes channel identifiers and conversion times, check whether the rerank mark and sorting logic of the retrieval results meet the configuration requirements.
- View the retrieval logs, confirm that cross-source hybrid recall covers core data sources such as advertising platforms and CRM, and the number of returned results matches the single retrieval limit setting.
- Test long-text attribution rule queries, confirm that the rerank model is called normally, and the result sorting combines core fields such as channel priority and conversion times.
The values given 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.