This document applies to FastGPT v4.17.0.
What this category of data looks like
Data for this category comes from ticketing systems, internal chat logs, product operation logs, and customer contract SLA documents. Daily processed documents include original customer-submitted tickets, pre-configured issue categorization tables, historical processing records, and service level agreement reports. Data updates occur in real time or daily. Each individual data entry contains fields such as ticket ID, submission time, issue description, impact scope, associated team, and SLA constraints. No additional statistical percentage data is included.
What constraints these characteristics impose on the reranking and hybrid retrieval link
Real-time or daily updates to ticket data require the retrieval pipeline to support incrementally synced data sources, to avoid matching outdated processing records. The mixed structured and unstructured structure of multi-field data requires hybrid retrieval to cover both unstructured issue descriptions and structured categorization rules and SLA constraints, to prevent missing critical information from single retrieval dimensions. Each data entry includes multiple pieces of information such as associated team and impact scope. The reranking stage must prioritize sorting tickets with high impact scope and approaching SLA deadlines. It must also support weight allocation across different data sources, to avoid irrelevant historical records interfering with current ticket acceptance workflows.
How to set configurations
| Configuration Item | Recommended Values | Rationale |
|---|---|---|
| Reranking Toggle | Enabled | Ticket issue categorization requires prioritizing matching semantically similar historical processing solutions. Reranking improves the sorting priority of highly relevant results |
| Similarity Threshold | 0.75–0.85 | Filter low-relevance historical tickets, retain reference data with high semantic matching to current issues, and adapt to detail differences in ticket descriptions |
| Single Retrieval Limit | Top 8–12 entries | Control the total number of retrieval results, balance retrieval latency and reference data coverage, and meet the real-time response requirements of operations on-duty workflows |
| Retrieval Mode | Hybrid Retrieval | Cover both structured categorization rules and unstructured historical ticket data simultaneously, meeting the dual needs of rule matching and experience reference in ticket acceptance |
| Reranking Model Call Limit | Top 3–5 entries | Limit the number of results processed in the reranking stage, reduce resource usage of the reranking model, and avoid GPU memory overload |
Three common misconfigurations
- Hybrid retrieval result response time exceeds 10 seconds, and single retrieval limit is set above 15 entries. Troubleshooting: Check the single retrieval limit configuration, confirm whether unconstrained retrieval result quantity has caused pipeline latency to exceed real-time response requirements.
- GPU memory usage grows rapidly to 6–7 GB after reranking model calls, with multiple GPUs occupied simultaneously. Troubleshooting: Check the reranking model call limit configuration, confirm whether the number of reranking results per request is not limited, or the concurrent call limit for the reranking model is not configured, causing multiple requests to occupy GPU resources simultaneously.
- Rerank error is prompted when connecting to a reranker model offline. Troubleshooting: 1. Check whether the FastGPT server has configured the
AIPROXY_API_ENDPOINT(root address of the AI Proxy service, for example http://aiproxy:3000, do not append paths such as /v1 or /chat/completions) andAIPROXY_API_TOKEN(AI Proxy administrator token, must match the AI Proxy server configuration and cannot be empty) environment variables; 2. Check whether relevant information for the reranking model has been correctly filled in the model channel configuration, confirm that the model call pipeline configuration is correct.
How to confirm successful configuration
- Check the FastGPT service startup logs, confirm there are no startup validation errors for missing AI Proxy environment variables.
- Submit a test ticket, check whether the sorting of retrieval results prioritizes historical records with high impact scope and approaching SLA deadlines.
- Monitor the resource usage of the reranking model, confirm that single GPU memory does not experience abnormally rapid growth.
- Compare the latency differences between hybrid retrieval and pure vector retrieval, confirm that hybrid retrieval latency meets business response requirements.
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.