What the data for this category looks like
This category of data mainly comes from business systems, App permission lists, privacy policies, and user request records. The update cycle is quarterly or when a single product change occurs. Daily processed documents include personal information lists, data processing rules, impact assessment reports, and user authorization texts. Most document fields include information subject identifiers, processing purposes, storage periods, compliance bases, etc. The length of individual documents varies widely, ranging from hundreds to thousands of characters.
What constraints these characteristics impose on the reranking and hybrid retrieval link
Since data sources are scattered and document structures vary greatly, hybrid retrieval must match both structured fields and unstructured text content, and balance the weights of different documents. Update frequency is not fixed, and a single change may affect multiple documents, so retrieval results must prioritize the latest updated compliance documents. Documents revolve around information that can identify natural persons, so retrieval results must strictly filter irrelevant content to avoid leakage or misuse of personal information. Meanwhile, compliance scenarios have high accuracy requirements for retrieval results, and the relevance of reranked results must match compliance requirements consistently, avoiding the introduction of non-compliant reference materials.
Configuration Settings
These settings apply to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled | Compliance scenarios require highly relevant results; reranking improves retrieval accuracy and filters low-match content |
similarity | 0.75–0.85 | Filter low-relevance content, avoid introducing irrelevant non-compliant documents, and ensure compliance relevance of retrieval results |
limit | Top 8–12 entries | Balance retrieval efficiency and result coverage, meet the batch verification needs of compliance documents, and avoid excessive results increasing processing burden |
searchMode | Hybrid retrieval mode | Match both structured fields and unstructured compliance text, covering multi-source compliance data |
| Single search return limit | 20 entries | Control the scale of retrieval results, avoid subsequent processing timeouts caused by excessive compliance materials |
Three Common Misconfigurations
- Phenomenon: Retrieval response times out after enabling
usingReRank, returning status code 504. Cause: The single document length of compliance documents is not adapted, and the reranking model's loading or inference time exceeds the system's default timeout threshold. - Phenomenon: The number of reranked knowledge base search results does not match the configured
limit, and the actual returned number is insufficient. Cause: In hybrid retrieval mode, the initial recall results do not reach the minimum number required for reranking, and the reranking logic automatically truncates the results without returning the specified number as configured. - Phenomenon: Low-relevance non-compliant documents still appear in retrieval results after configuring the
similaritythreshold. Cause: Structured field matching is not enabled at the same time, and only unstructured similarity calculation is relied on, which cannot filter irrelevant field content, resulting in low-relevance documents being included in the result set.
How to Verify Proper Configuration
- Navigate to the FastGPT knowledge base configuration page, check the
usingReRankswitch status, and confirm it is enabled. - Submit a compliance-related test query, review the sorting logic of retrieval results, and confirm that highly relevant compliance documents are ranked first.
- Adjust the
similaritythreshold, verify the filtering effect of retrieval results, and confirm that low-relevance content is excluded. - Check system logs to confirm that the reranking model's call duration is within a reasonable range, with no timeout errors.
The values provided here are common starting points and should be evaluated 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.