This document applies to FastGPT v4.17.0 and later versions.
What the data for this category looks like
Data for this category primarily comes from advertising delivery platforms, creative material libraries, user interaction data, and conversion data. Update frequencies are mostly daily or weekly. Regular associated documents include creative scripts, material test sheets, delivery data reports, and iteration checklists. Core fields include material type, copy content, delivery time slots, click volume, conversion volume, and more. There are no unified fixed units, and adjustments are made based on specific test objectives.
What constraints do these characteristics impose on the reranking and hybrid retrieval link
Data for this category is scattered across multiple sources such as advertising platforms and material libraries. Hybrid retrieval must recall creative materials and delivery data from different systems simultaneously. Clear association rules for cross-source data must be defined. The daily or weekly update rhythm requires the retrieval pipeline to adapt to batch data processing, avoiding overload from single-round recall. The core goal of creative testing is to screen high-conversion combinations. The reranking link must prioritize sorting based on business indicators such as clicks and conversions, while supporting material metadata in different formats, to avoid screening bias caused by relying solely on semantic similarity.
How to configure settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Similarity threshold | 0.75–0.85 | Filters low-relevance creative materials, adapting to the semantic matching accuracy requirements of information feed ad creatives |
| Single return limit | Top 10–15 items | Controls the number of candidate creatives returned, avoiding exceeding workflow context window limits |
| Search mode | Hybrid retrieval | Matches the recall requirements for multi-source data in this category, integrating semantic and keyword retrieval results |
| Enable reranking | Enabled | Performs secondary sorting on recall results based on business indicators such as clicks and conversions, raising the priority of high-potential creatives |
| Reranking model matching degree | Calibrated via actual testing | Adapts to the feature distribution of materials in this category, avoiding incorrect screening of high-quality creative combinations |
Three common mistakes
- After configuring a custom rerouting channel, calls fail with an interface authentication error prompt. Troubleshooting: If the credential field for the custom rerouting channel is filled with the AI Proxy administrator Token, replace it with the supplier credential configured for the upstream model channel.
- Invalid citations cannot be filtered in hybrid retrieval scenarios. Troubleshooting: Check whether the retrieval configuration enables citation filtering rules, or if metadata filtering fields for the corresponding data source are bound.
- Multiple variable outputs in workflows cannot be associated with a specified knowledge base for reranking. Troubleshooting: Verify the reranking link has variable mapping rules configured to match creative materials from the corresponding data source.
How to confirm the configuration is complete
- Verify the
AIPROXY_API_ENDPOINT(root address of the AI Proxy service) andAIPROXY_API_TOKEN(AI Proxy administrator Token) environment variables for the App/Admin server are correctly configured, with no appended paths such as /v1 and non-empty values. - Trigger a creative test retrieval task, and check that the number of returned results matches the configured range for the single return limit.
- Compare the sorting of retrieval results with reranking enabled and disabled, confirming creatives with high conversion indicators are displayed first.
- View workflow logs to confirm no timeout or authentication failure error messages appear in the reranking link.
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.