What data for this category looks like
Data for this category originates from customer data center environments, equipment inventories, software packages, and network topologies. Daily deliverables include deployment manuals, installation records, configuration checklists, on-site sign-in sheets, and related documents. Data is updated per order or project phase. Core fields include device model, IP address, software version, configuration parameters, and on-site operation steps. No universally applicable fixed dataset exists. All data only covers exclusive information for a single on-site deployment.
Constraints on reranking and hybrid retrieval from these characteristics
On-site deployment datasets are exclusive to single projects and updated frequently. Reranking and hybrid retrieval must prioritize the current customer’s on-site information, and avoid recalling irrelevant general deployment content. Most data fields are precise device parameters and step descriptions, so retrieval must balance semantic matching and exact matching. On-site troubleshooting requires fast response times, so the number of retrieval results must be kept within a reasonable range to avoid excessive redundant information interfering with engineer judgment. Data is only valid during the project cycle, so retrieval scope must be limited to documents related to the current deployment to prevent cross-project data confusion.
How to configure settings
These configurations apply to FastGPT v4.17.0 and later versions:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Rerank Toggle | Enabled | On-site deployment datasets are exclusive to single projects. Reranking filters out general deployment documents and focuses on the current customer’s device and configuration information |
| Single Recall Limit | Top 5–10 results | On-site deployment troubleshooting typically involves a small number of core steps and devices. Excessive results increase engineer filtering costs |
| Similarity Threshold | 0.75–0.85 | Matches exact fields such as customer on-site device models and network parameters, avoiding recall of irrelevant general cases |
| Retrieval Mode | Hybrid Retrieval | Combines semantic matching from vector retrieval and keyword matching, covering step descriptions in deployment manuals and exact fields in configuration checklists |
| Retrieval Timeout | 60 seconds | On-site troubleshooting requires fast response times. Excessive timeout delays on-site support progress |
Three common configuration errors
- Rerank results appear empty or are not applied in the conversation details page. Troubleshooting: Server environment variables
AIPROXY_API_ENDPOINT(root address of the AI Proxy service) andAIPROXY_API_TOKEN(AI Proxy admin token) are not properly configured in the FastGPT App/Admin server, orAIPROXY_API_ENDPOINTis set to a single model address instead of the root address of the AI Proxy service. - The number of retrieval results greatly exceeds the configured single recall limit. Cause: The hybrid retrieval recall logic is not correctly bound to post-reranking result filtering, causing the first-recall-then-filter step to fail.
- A model connection timeout is returned after invocation. Troubleshooting: The AI Proxy admin token was entered into the application API authentication field, or the locally deployed rerank service cannot be accessed normally by the FastGPT service.
How to verify correct configuration
- Check the FastGPT service startup logs to confirm there are no startup validation errors for missing
AIPROXY_API_ENDPOINT(root address of the AI Proxy service) orAIPROXY_API_TOKEN(AI Proxy admin token) server environment variables. - Initiate a test retrieval targeting the current project’s deployment documents, and confirm that retrieval results prioritize displaying the current customer’s device models and configuration parameters.
- View the retrieval stage logs in the conversation details page, and confirm that the reranking step was executed and the number of results matches the configured single recall limit.
- Verify connectivity of the local rerank service, ensuring the FastGPT service can send normal requests to the rerank model’s interface.
The values listed 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.