What the data for this category looks like
Daily documents for this category include onboarding application forms, tenant configuration sheets, administrator account lists, and delivery confirmation forms. Data sources are CRM systems, contract order databases, identity management systems, and product backends. Data is updated per order or daily, and the number of documents associated with a single work order is relatively fixed. Most fields include customer identifiers, tenant IDs, permission configuration items, onboarding status, and similar attributes. Typical processing entities rely on these documents to complete customer information verification, permission initialization, and onboarding result confirmation. Errors in this process will prevent customers from logging in or block delivery workflows.
What constraints these characteristics impose on re-ranking and hybrid retrieval
Retrieval needs for this category mostly revolve around troubleshooting exceptions in the tenant onboarding process. Therefore, the retrieval scope must be strictly limited to documents associated with the current customer to avoid cross-tenant data interference. Since data is updated per work order unit, the dataset size for a single retrieval is small, so no large retrieval upper limit is required. Additionally, documents include structured configuration sheets and unstructured delivery confirmation forms, so capabilities combining full-text retrieval and structured field matching are needed to locate both configuration parameters and process steps. Furthermore, retrieval results must prioritize matching key stages of the onboarding process, so the re-ranking stage must assign higher weight to content related to the onboarding workflow.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
usingReRank | Enabled | Most inquiries for this category are process-related exceptions; re-ranking can prioritize matching key steps of tenant onboarding to improve troubleshooting efficiency |
limit | Top 3–5 results | The number of documents associated with a single work order is limited; limiting results to a small number reduces irrelevant information and speeds up response times |
similarity | 0.75–0.85 | Filters low-match irrelevant documents to avoid interference from configuration information of non-current tenants |
searchMode | Hybrid retrieval | Data for this category includes structured tenant configuration sheets and unstructured delivery documents; hybrid retrieval balances full-text and field matching capabilities |
Three Common Misconfigurations
- Symptom: Re-ranking markers are empty in retrieval results, and returned results are not sorted by matching priority. Cause:
usingReRankis not enabled in the retrieval configuration, or the re-ranking model is not bound to the current knowledge base’s retrieval chain. - Symptom: Re-ranking works normally in knowledge base test queries, but fails when invoked via large model chat. Cause: Re-ranking is not specified as enabled during large model calls, or re-ranking model configuration is not synchronized to the large model’s channel settings.
- Symptom: Only re-ranking models appear in the channel model list, and other language models cannot be selected. Cause: Global configuration for multi-model channels was not completed during commercial deployment, and only a dedicated call chain for the re-ranking model was bound.
How to Confirm Configuration Is Complete
- Navigate to the knowledge base retrieval test page in FastGPT v4.17.0, enter queries such as "tenant initialization failed" or "administrator account cannot log in", and check if retrieval results have re-ranked sorting markers.
- Invoke a large model bound to this knowledge base to initiate a query, and check if returned results prioritize documents covering key steps of the tenant onboarding process.
- Access the retrieval configuration page and confirm that the
usingReRanktoggle is enabled, and the correct re-ranking model is bound. - Review the channel model configuration list and confirm that other language models besides the re-ranking model have been added and linked to the current retrieval chain.
Note: The provided values are common starting points and should be evaluated against the reader’s own sample datasets.
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.