What this category of data looks like
This category of data primarily comes from advertising platform backends, attribution analysis systems, CRM lead databases, and landing page data collection modules. Daily processed documents include daily campaign performance reports, weekly budget review reports, price adjustment records for individual creatives, multi-round creative test comparison tables, and optimization strategy review notes. Data updates mostly occur via daily batch synchronization, while real-time data for core campaign metrics is pushed at minute-level intervals. Document fields include impressions, clicks, conversions, cost per acquisition, creative click-through rate, and more, with units mostly being counts, yuan, and percentage.
What constraints these characteristics impose on reranking and hybrid retrieval
Data sources for this category are scattered and diverse in type. Hybrid retrieval needs to recall data from advertising platform backends, attribution analysis records, and internal business documents simultaneously, so recall priorities for different data sources must be clearly defined. Data update rhythms include two categories: daily batch synchronization and minute-level real-time pushing. The retrieval scope must support filtering by time interval to avoid including expired historical data. Daily documents vary widely in structure, ranging from short individual price adjustment records to long weekly review reports. Reranking models need to adapt to input content of different lengths. Abnormal core campaign metric positioning requires quickly matching highly relevant content, so the sorting logic for retrieval results must prioritize fields related to campaign performance to improve the accuracy of abnormal metric localization.
How to set configurations
This document is written based on FastGPT v4.17.0. All following configuration items use official specified parameters.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled, automatically triggers for retrieval requests where input text length ≥ 800 characters | Adapts to the varied document lengths of this category, avoids wasting resources on short texts by skipping reranking, and improves sorting accuracy for long texts via reranking |
similarity | Value range of 0.65–0.80 | Matches the semantic relevance of core campaign metrics in this category, avoids recalling irrelevant historical records or low-relevance creative test data |
limit | Top 10 recalled results | Controls the number of results returned per retrieval, adapts to the rapid positioning needs of operational on-call and ticket Q&A, and avoids excessive results interfering with abnormal metric troubleshooting |
searchMode | Hybrid recall mode, integrates advertising platform data and internal business documents | Adapts to the retrieval needs of multi-source data for this category, covering different types of content such as real-time advertising platform data and internal review notes |
Three common misconfigurations
- Symptom: Custom rerank model API address is filled correctly, but retrieval results are not reranked, and sorting matches the initial recall results. Cause: The
usingReRankconfiguration item is not enabled, or the text length threshold for triggering reranking is set too high, failing to match the actual length of the current retrieval content. - Symptom: When using the Qwen3-Reranker-8B model, the interface shows model loading successful, but retrieval returns rerank result as false. Cause: The model’s input and output format is not correctly mapped in the model configuration, causing FastGPT to fail to parse reranked results.
- Symptom: When retrieving long content across multiple pages, the rerank model is not called, and only the top N initial recall results are returned. Cause: Rerank trigger rules are not configured for long text scenarios; FastGPT defaults to triggering reranking only for short texts.
How to confirm configurations are correctly set
- View FastGPT retrieval logs to confirm rerank model call records, and check for successful requests and responses.
- Manually input a query containing campaign metrics, and compare the sorting differences of retrieval results between enabling and disabling the rerank function.
- Adjust the length of retrieval content to verify whether the rerank model is triggered when the text length reaches the set threshold.
- Check the custom rerank model API configuration to confirm that the request address and parameter format match the model deployment requirements.
The values provided here 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.