What This Category of Data Looks Like
Data for this category originates primarily from internal enterprise databases, data dictionaries, data masking tool operation logs, and access approval systems. Daily maintained documents include masking plans, field mapping tables, masking verification reports, and approval records. Data is updated per project batch or single task. Data fields include sensitive field identifiers, masking rule codes, verification results, approval status, and more. No unified fixed units exist; values adjust flexibly based on project-specific configurations.
Constraints Imposed by These Characteristics on Reranking and Hybrid Retrieval
Data for this category is updated per project batch, and datasets from different projects are independent. Reranking and hybrid retrieval must support isolating retrieval scopes by project dimension to avoid cross-project sensitive data mixing. Document types include structured field mapping tables, long-text masking plans, and reports with verification results. Retrieval must support matching logic for multi-format data. Since data includes sensitive field identifiers and verification results, the reranking link must first filter or desensitize non-authorized visible retrieval results, while ensuring compliance-related queries are accurately recalled. Data updates have no fixed cycle, so retrieval configurations must support dynamically adjusting the recalled dataset scope to adapt to different project iteration rhythms.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Enable Reranking | Enabled | This category of data contains sensitive fields. Reranking is required to filter low-correlation results and avoid retrieving irrelevant sensitive data |
| Similarity Threshold | 0.75–0.85 | Balances retrieval accuracy and compliance for sensitive data. Avoids retrieving irrelevant data with too low a threshold, and missing compliance-related content with too high a threshold |
| Single Return Limit | Top 5–8 entries | Single-project data volume for this category is moderate. Limiting return entries reduces sensitive data leakage risks while controlling retrieval response time |
| Search Mode | Hybrid Retrieval | Requires matching both structured field mapping tables and unstructured masking plan documents. Hybrid retrieval balances exact matching and semantic matching |
| Reranking Model Call Timeout | 30–60 seconds | Adapts to the length differences of documents in this category, avoiding task failure caused by retrieval timeouts for long-text masking plans |
| Reranking Model Selection | Determined via actual testing | Must adapt to the FastGPT v4.17.0 environment and the semantic characteristics of this category’s data, avoiding retrieval result abnormalities caused by model incompatibility |
Three Common Configuration Errors
- Symptom: Retrieval response time exceeds 60 seconds after enabling reranking. Cause: Single return limit is not configured, or reranking model call timeout is set too short, leading to large volumes of data entering the reranking link.
- Symptom: Knowledge base search result reranking does not take effect, and returned results match those when reranking is disabled. Cause: Search mode is not correctly set to hybrid retrieval, or reranking model deployment adaptation is incomplete.
- Symptom: Reranking models deployed via Docker call successfully, but returned results are always false. Cause: Reranking model input format does not match FastGPT interface requirements, or sensitive field filtering logic incorrectly marks valid results as invalid.
How to Verify Correct Configuration
- Navigate to the FastGPT knowledge base configuration page, confirm that the enable reranking switch is turned on, and search mode is set to hybrid retrieval.
- Initiate a query targeting masking plans, verify whether the returned result ranking differs from the initial recalled results.
- Check retrieval logs, confirm that reranking model call duration matches the configured timeout threshold, with no timeout errors.
- Submit a query containing sensitive field keywords, verify that non-authorized visible content has been filtered from returned results.
Important Note
The configuration values listed here are common starting points. Validate them against your own specific data samples and testing needs.
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.