What data in this category looks like
This category of data primarily comes from raw datasets, ETL operation logs, annotation platform output results, and sampling quality inspection forms. Daily generated documents include cleaning rule configurations, abnormal data lists, quality inspection sampling reports, and delivery result descriptions. Data is updated daily or per batch task cycle. Fields include data identifiers, anomaly types, processing status, sampling qualification rates, and other related fields. No unified fixed units are used, and units adjust based on the dimensions of the specific dataset.
What constraints these characteristics impose on the reranking and hybrid retrieval link
Raw dataset scale fluctuates with business batches. ETL logs and sampling results require time-based associated retrieval. Hybrid retrieval must balance structured fields and unstructured text. Cleaning rule documents and abnormal lists have high update frequencies. Retrieval results must cover the latest configurations and abnormal records. Data fields have diverse dimensions. Reranking must filter low-confidence association results to avoid mixing mismatched abnormal data with cleaning rules. Sampling quality inspection results must prioritize matching the current batch’s task identifiers to prevent cross-batch data from interfering with retrieval accuracy.
How to set configurations
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
similarity | 0.72–0.88 | Matches the field association strength of this category of data, filters low-relevance cross-batch data |
limit | Top 15–25 entries | Adapts to the scale of single-batch data, avoids returning excessive redundant abnormal data entries |
searchMode | hybrid | Balances structured task identifier fields and unstructured anomaly description text |
usingReRank | Enabled | Reorders recalled results by anomaly severity and association timeliness, prioritizes displaying high-priority anomalies |
Three common configuration errors
- Symptom: Reranking model tests return errors, and the interface prompts model loading failure. Cause: The deployment address of the private reranking model is not configured to the corresponding retrieval node, or the model version does not meet the adaptation requirements of FastGPT v4.17.0.
- Symptom: After enabling hybrid retrieval, the number of returned results does not match the set
limit, and entries exceeding the upper limit appear. Cause: Non-essential recall data sources are not properly closed, or thelimitconfiguration is not applied to all links of the hybrid retrieval process. - Symptom: When reranking is not enabled, the relevance field of retrieval results has no clear reference value. Cause: Weighting is not combined with the structured fields of this category of data. Basic embedding matching relevance only covers text semantics, and does not associate key dimensions such as task batches and anomaly types.
How to confirm configurations are properly applied
- The retrieval configuration page is checked to confirm the
usingReRankswitch is enabled andsearchModeis set to hybrid. - A batch of test data is submitted, the abnormal data list is retrieved, and the returned results are checked to confirm they are sorted from highest to lowest anomaly severity.
- The value range of
similarityis adjusted to verify whether the relevance of retrieval results changes with configuration adjustments. - The number of entries returned by retrieval is checked to confirm it meets the
limitsetting requirements, and no additional redundant entries are mixed in.
The values provided are common starting points and should be validated against local test 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.