What the data for this category looks like
This category's data primarily comes from data asset inventories, database metadata, business system descriptions, and compliance regulatory documents. Daily produced documents include data catalogs, classification and grading rules, important data identification tables, and review minutes. Data is updated quarterly or synchronized after each business system change. Document fields include asset unique identifiers, sensitive data types, security classification levels, affiliated business domains, associated review records, and more. There are no unified fixed units. Some fields exist as text enumerations.
Constraints Imposed on Reranking and Hybrid Retrieval
Data for this category originates from scattered sources, including structured metadata and unstructured documents. Hybrid retrieval must cover both structured asset identifiers and unstructured rule text. Data updates occur slowly, but changes require rapid synchronization to the retrieval database. The reranking link must therefore support filtering recall results by update time. Most document fields are enumerated text. Reranking weights must match the business priority of fields. For example, the sensitive data type field has a higher weight than the affiliated business domain field. The number of data assets grows gradually as business systems expand. The retrieval recall upper limit must adapt to changes in asset scale to avoid returning too many low-relevance entries.
How to Configure
The following configurations apply to FastGPT v4.17.0 and above.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
usingReRank | Enabled | This category's data includes sensitive field enumerations and rule text. Reranking improves retrieval ranking accuracy for high-priority content |
similarity | 0.72–0.85 | Text similarity threshold for matching classification and grading rules. Prevents missed or incorrect identification of sensitive data types |
limit | Top 10–15 entries | Adapts to the retrieval scale of single-batch data assets. Balances recall completeness and interface response latency |
searchMode | Hybrid retrieval mode | Covers both structured metadata and unstructured documents. Meets retrieval requirements for multi-source data |
| Reranking Recall Upper Limit | Top 3–5 entries | Focuses on high-weight sensitive fields and rule entries in this category. Reduces interference from invalid results |
Three Common Misconfigurations
- Phenomenon: Validation fails when starting the App/Admin service, and a prompt indicates missing required environment variables. Cause:
AIPROXY_API_ENDPOINT(root address of the AI Proxy service) andAIPROXY_API_TOKEN(AI Proxy administrator token) are not configured properly, or the entered content is not an AI Proxy administrator token, such as model vendor keys or application API Keys. - Phenomenon: An error occurs after configuring a local reranking model and running the service, with a prompt indicating model loading failure. Cause: The deployment address of the reranking model is not correctly entered in the model channel configuration, or the local reranking model service is not started.
- Phenomenon: The number of retrieval results does not match the configured limit, with too many or too few recall entries. Cause: The hybrid retrieval mode is not correctly associated with structured and unstructured data sources, leading to abnormal retrieval scope.
How to Verify Successful Configuration
- Check the App/Admin service startup logs to confirm there are no missing environment variable errors related to AIPROXY.
- Launch an internal ticket Q&A test, and review whether retrieval results are sorted by sensitive field priority to verify that the reranking logic is active.
- Adjust the similarity threshold, observe changes in the number of retrieval results to confirm that the configuration parameters take effect.
- Verify the model channel configuration, and confirm that the reranking model deployment status is normal.
The provided values are common starting points and should be evaluated 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.