What the data for this category looks like
This category of data comes primarily from risk control systems, audit records, user behavior logs, and report evidence. Daily processing involves documents including account ban records, appeal materials, review opinions, and user notifications. Each document contains fields such as unique account identifier, violation type, disposal time, appeal submission time, and review result. Data updates occur daily or in real time, synchronized with user appeal submissions and audit process progress.
What constraints these characteristics impose on the reranking and hybrid retrieval workflow
Decentralized data sources require retrieval across multiple system documents, and hybrid retrieval of ban records and appeal materials in varying formats. Real-time updated data requires retrieval links to support low-latency responses, to avoid delaying appeal review timelines. Account-level associated data demands precise matching. The reranking step prioritizes associating historical violations and appeal records for the same account, to prevent cross-account matching errors. Field differences across documents require unified mapping of keyword dimensions during retrieval, to ensure recall accuracy for hybrid retrieval and avoid ineffective retrieval caused by field mismatches.
How to set configurations
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| similarity threshold | 0.63–0.77 | Matches semantic relevance between appeal content and historical violation records, and filters low-correlation redundant documents |
| single recall limit | 10–14 entries | Limits the total number of documents returned by a single hybrid retrieval, adapting to the processing pace of customer service rapid verification |
| hybrid retrieval mode | Enabled | Integrates vector retrieval and keyword retrieval, covering retrieval needs for structured ban records and unstructured appeal materials |
| rerank toggle | Enabled | Reranks recalled documents by relevance to the current appeal, prioritizing display of higher-correlation review reference materials |
| rerank return count | Top 6 entries | Limits the number of documents output after reranking, controlling the information load of the large model context |
This document applies to FastGPT v4.17.0.
Three common misconfigurations
- Phenomenon: The rerank function works normally during independent knowledge base testing, but rerank results always return false when integrated with large model calls. Cause: The rerank trigger condition was not configured in the large model call link, and the rerank toggle was only enabled in the knowledge base testing phase.
- Phenomenon: Hybrid retrieval only returns keyword-matched documents, and vector retrieval recall results are not included in the retrieval scope. Cause: Hybrid retrieval mode was not enabled correctly, and only a single keyword retrieval or vector retrieval method was activated.
- Phenomenon: After commercial edition deployment, the channel model configuration interface only allows selection of rerank models, and no other large language models can be added. Cause: Permission binding configuration for channel models was not completed, and the association logic between rerank models and large language models was not properly initialized.
How to confirm configurations are correct
- The knowledge base retrieval test page is accessed, appeal-related keywords are entered, and returned documents are verified to be sorted by relevance.
- Retrieval link operation logs are reviewed, and hybrid retrieval mode and rerank toggle are confirmed to be enabled.
- A test target account ID is configured, historical ban and appeal records for that account are retrieved, and retrieval results are verified to accurately associate with the target account.
- A large model is integrated to initiate an appeal review request, and returned reference documents are checked to meet the preset rerank return count requirement.
The configuration values provided are common starting points, and should be evaluated against local 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.