What this type of data looks like
This data primarily originates from customer historical business systems, CSV/Excel templates, database export files, and product import logs. Daily supporting documents include data import templates, migration implementation plans, field mapping comparison tables, and post-import verification reports. The data update schedule is per order or irregular, adjusted based on customer migration requirements. Data fields include business entity identifiers, business attribute values, and association relationship identifiers. Units include data row counts, storage capacity, business timestamps, and more. Format alignment must be completed through field mapping.
What constraints these characteristics impose on re-ranking and hybrid retrieval
Multi-source heterogeneous data sources create differences in vector representations across different data formats. Hybrid retrieval must adapt to recall logic for structured database exports, semi-structured CSV templates, and unstructured import logs simultaneously. The data update schedule is not fixed, and field mapping rules differ for each migration. The matching threshold of the re-ranking model must be adjusted for each individual migration. Supporting field mapping tables and verification reports must be embedded in the retrieval context. The re-ranking link must prioritize matching aligned business fields to avoid misaligned recalls. Single migration data volume varies widely. The number of returned entries for retrieval and re-ranking must be limited to prevent resource overload.
The following configurations apply to FastGPT v4.17.0.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.65–0.75 | Aligns with the field mapping matching accuracy of customer imported data, filters low-correlation misaligned field results |
limit | Top 10 entries | Controls the number of candidate results returned per single retrieval to avoid resource overload during the re-ranking link |
searchMode | hybrid | Adapts to multi-source heterogeneous imported data, mixes structured query and unstructured semantic recall logic |
usingReRank | Enabled | Corrects the relevance ranking bias of basic retrieval for the field association characteristics of imported data |
| Re-ranking model matching threshold | Calibrated via actual testing | Adapts to business semantic differences of different customer migration data, avoids incorrectly filtering valid matching results |
Three common configuration errors
- Phenomenon: Retrieval returns results that are not reordered as expected, and the number of returned entries does not match the configured
limit. Cause: TheusingReRankconfiguration item is not enabled correctly, or the re-ranking model call times out and the link is skipped. - Phenomenon: Hybrid retrieval returns a large number of low-correlation misaligned field results. Cause: The
similaritythreshold is set too low, failing to filter mapped fields with insufficient matching degrees. - Phenomenon: The re-ranking link has no response or reports an error directly. Cause: The corresponding re-ranking model interface is not bound on the FastGPT platform, or the
limitvalue is too large, causing insufficient model resources.
How to verify successful configuration
- Navigate to the retrieval configuration page on the FastGPT platform, confirm that the
usingReRankswitch is in the enabled state, andsearchModeis set to hybrid. - Import a set of test migration data, execute retrieval, and check the sorting logic of returned results to verify whether re-ranking takes effect.
- Adjust the
similaritythreshold, observe changes in the number of returned results, and confirm that the filtering effect of the threshold configuration on retrieval results meets expectations. - Check import logs and retrieval logs, confirm that there are call records for the re-ranking link and no error messages.
The values provided are common starting points and should be validated against local deployment 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.