What the data for this category looks like
The core data for this category comes from four sources: advertising platforms, user profile databases, budget scheduling tables, and attribution tools. Daily processed documents include four types: delivery plans, targeting settings, budget schedules, and conversion tracking configurations. Fields cover targeted audience tags, budget amounts, conversion event IDs, delivery time periods, and more. Units include yuan, visits, days, etc. The data update frequency is daily or triggered by individual delivery activities. Adjustments to some configuration items are immediately synchronized to associated data sources.
What constraints these characteristics impose on the reranking and hybrid retrieval link
Data sources are scattered and include both structured configuration documents and unstructured records, requiring hybrid retrieval to cover the recall scope of both types of data. The data update frequency is high, so retrieval results must match the latest configuration adjustments to avoid returning outdated delivery plan content. Most document fields are business-specific configuration items; basic embedding retrieval can easily confuse similar but functionally different configurations, so reranking is needed to filter irrelevant results. Additionally, the core time-consuming links for this category focus on audience targeting and budget allocation. Reranking must prioritize retrieval results for these high-value configuration items to reduce interference from irrelevant information.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled | Documents for campaign setup include multiple types of business-specific configurations. Enabling reranking filters irrelevant results returned by embedding retrieval and improves the accuracy of configuration references |
limit | Top 8 entries | There are four core daily document types. Recalling the top 8 entries covers major configuration scenarios and avoids excessive redundant information interfering with the efficiency of technical support and pre-sales Q&A |
similarity | 0.72–0.85 | The semantic similarity of configuration documents for this category has high differentiation. A threshold lower than 0.72 will introduce irrelevant delivery effect optimization content, while a threshold higher than 0.85 will miss valid configuration references |
searchMode | Hybrid Retrieval | It is necessary to recall both structured configuration documents from advertising platforms and unstructured records from budget tables and attribution tools. Hybrid retrieval balances the recall effects of both types of data |
Three common mistakes
- Symptom: Query speed slows significantly after configuring
usingReRank. Cause: A lightweight rerank model compatible with FastGPT v4.17.0 is not used, and a model exceeding resource limits is loaded, causing service lag. - Symptom: No change in retrieval results after enabling result reranking. Cause:
searchModeis not set to hybrid retrieval, only basic embedding retrieval logic is triggered, and the rerank model is not called for secondary sorting. - Symptom: Docker-deployed reranker returns false. Cause: The model interface access address is not configured correctly, or the request does not carry valid authentication parameters, resulting in interface call failure.
How to confirm the configuration is complete
- View the retrieval logs of the FastGPT application, confirm that there are rerank model call records and valid sorted results are returned.
- Manually enter a query targeting audience configuration, compare the sorting differences of retrieval results before and after enabling the rerank function to verify that the rerank logic takes effect.
- Check the application configuration page, confirm that
usingReRankis set to enabled andsearchModeis set to hybrid retrieval. - Simulate common user queries related to budget scheduling, verify that retrieval results prioritize returning budget scheduling documents, which meets business priority 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.