What data for this category looks like
Data for this category primarily comes from object storage, data warehouses, annotation platforms, and customer acceptance feedback. Updates are triggered per batch or version. Daily produced documents include dataset descriptions, version records, delivery lists, and acceptance reports. Core fields include version number, dataset source identifier, total sample count, delivery time, acceptance status, and other standard fields. No non-standard units are used. The data structure is gradually supplemented with each version iteration.
What constraints these characteristics impose on reranking and hybrid retrieval
Data for this category is stored in batches by version. Datasets from different versions have differing samples. Retrieval must not mix version data. A single retrieval must link to the delivery list and acceptance records for the corresponding version. Therefore, reranking and hybrid retrieval must limit data source scope to avoid interference from cross-version samples. Data update frequency is high. Retrieval configurations must adapt to rapid switching of batch data. Version number fields filter invalid data to ensure pre-sales Q&A and technical support responses match the current delivery version. This prevents customer acceptance failures caused by inconsistent data versions.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Rerank Enable | Enabled | This category requires retrieval based on accurately delivered samples. Reranking filters low-relevance cross-version data |
| Similarity Threshold | 0.75–0.85 | Adapts to the semantic similarity distribution of dataset samples, prevents recall of irrelevant version samples |
| Single Recall Limit | Top 8–12 entries | Matches the sample count range of a single delivery list, controls the amount of information returned for pre-sales Q&A |
| Retrieval Mode | Hybrid Retrieval | Combines vector retrieval and keyword retrieval, covers content with different document structures such as dataset descriptions and version records, improves the coverage of pre-sales Q&A |
| Rerank Model Timeout | 600 seconds | Adapts to the loading and inference duration of local rerank models, prevents retrieval interruptions due to timeout |
| Version Range Restriction | Bind to the current delivery version number | Prevents mixing of cross-version data, ensures retrieved content matches the dataset currently being accepted by the customer |
This configuration applies to FastGPT v4.17.0.
Three common mistakes
- Phenomenon: After enabling the rerank function in the interface, no rerank results are returned in the online recall test. Cause: The dataset corresponding to the delivery version is not bound. Reranking only takes effect within the limited version scope.
- Phenomenon: After deploying a local rerank model, the interface displays an Invalid URL (POST /v1/rerank) error. Cause: The API address of the rerank model is not configured to the /v1/rerank format supported by FastGPT, or the model is not loaded correctly.
- Phenomenon: Retrieval results include dataset samples from old versions. Cause: The version range restriction configuration is not enabled, leading to recall of cross-version data.
How to confirm configurations are set correctly
- Navigate to the retrieval test page of the knowledge base, enter keywords for the current version's dataset, check if returned results include content from the current version's delivery list.
- After enabling the rerank function, compare the number of recall results and relevance ranking before and after enabling to confirm the rerank logic is active.
- Switch to datasets from other versions, verify that retrieval results automatically switch to samples from the corresponding version, with no cross-version data mixed in.
- Check system logs to confirm no errors in API calls for rerank requests, and that parameters match the current configuration.
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.