What this category of data looks like
This category of data primarily comes from code repositories, dependency lists, license texts, and contribution records. Core documents include LICENSE files, NOTICE files, dependency license lists, and copyright notices. All focus on code distribution and usage permissions. Updates occur with each version release or dependency change. Document structures vary widely. Some are short license clauses. Others are long bulk dependency compliance lists. Fields include license type, copyright holder, dependency name, and compliance check tags. There is no unified fixed-length format.
What constraints these characteristics impose on the reranking and hybrid retrieval link
Structural differences and update patterns of this category of data create multiple constraints for the reranking and hybrid retrieval workflow. Short license clauses and long dependency lists coexist. Hybrid retrieval must balance recall accuracy across different text lengths. The high-frequency update feature requires the retrieval pipeline to quickly adapt to the latest compliance documents. It avoids returning expired compliance information. Core fields have strong professional attributes. The reranking process must prioritize matching key compliance keywords such as license type and copyright ownership. It improves the compliance relevance of results. Retrieval result errors related to compliance may trigger commercial risks. Therefore, retrieval accuracy takes priority over recall breadth.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled | This category requires accurate matching of compliance-related keywords. Reranking filters low-relevance non-compliant documents and elevates result priority |
limit | Top 10 entries | Compliance documents must cover core dependencies and core license items. Excessive results increase retrieval latency. Top 10 entries balance coverage and performance |
similarity | 0.75–0.85 | Balances recall of compliance-related content and avoidance of missing niche licenses or marginal compliance items. Calibrated via testing on FastGPT v4.17.0 |
searchMode | Hybrid retrieval mode | This category of data includes short license texts and long dependency lists. Hybrid retrieval balances recall accuracy across different text lengths |
Three common mistakes
- After enabling
usingReRank, the knowledge base returns results that match the base retrieval order, with no obvious adjustment. The cause is incomplete API binding configuration for the reranking model, or incompatibility between the FastGPT version and the reranking model version. - Single retrieval latency exceeds the business acceptable threshold, and timeout prompts appear on the page. The cause is an excessively high
limitvalue, or batch processing parameters of the reranking model not adapted to the long-text compliance documents of this category. - Retrieval results miss core compliance items, such as license information for key dependencies not being recalled. The cause is an excessively high
similaritythreshold, which filters some documents with low similarity but containing key compliance information.
How to confirm correct configuration
- View retrieval logs in the FastGPT backend. Confirm that reranking model call records exist and have no error messages.
- Enter a compliance-related query term. Compare the result order between base retrieval and reranked results. Confirm that highly relevant compliance documents appear at the top.
- Adjust the
similaritythreshold for multiple tests. Confirm that the range of recalled documents meets the coverage requirements of compliance checks. - Count single retrieval latency. Adjust the
limitvalue or reranking model configurations to bring latency in line with business delay requirements.
The values given 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.