What This Category’s Data Looks Like
This category of data primarily comes from customer interview records, public bidding documents, partner manufacturer product materials, and pre-sales technical solution documents from historical projects. Update frequency is adjusted per individual project or weekly. Each document includes four core content types: technical proposal, customer demand survey form, quotation list, and demo script. Fields cover technical routes, implementation boundaries, quotation configuration items, and clearly stated customer demand points. There is no unified standardized field format, so adaptation to the customized needs of individual projects is required.
What Constraints These Characteristics Impose on Reranking and Hybrid Retrieval
Data sources for this category are scattered and have inconsistent formats. Hybrid retrieval must accommodate structural differences across documents to avoid recall bias caused by inconsistent formatting. Individual project data is updated per project, so the retrieval pipeline needs to support real-time or near-real-time incremental updates. This ensures retrieval results match the latest needs of the current project. Customized document content means general recall rules cannot cover all scenarios. The reranking link must be used to prioritize matching the specific demand points raised by the current customer. Pre-sales Q&A scenarios require fast response times. Total time for retrieval and reranking must stay within a reasonable range to avoid disrupting customer communication rhythms.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.65–0.75 | Match the precision of technical details and customer demands in pre-sales proposals, and avoid recalling irrelevant general proposal content |
limit | Top 10 entries | Control the number of initial recall candidate results, balance retrieval efficiency and result coverage, and adapt to the limited demand pool of individual pre-sales projects |
searchMode | hybrid | Combine the advantages of keyword retrieval and semantic retrieval, adapt to scenarios where pre-sales proposals have both clear keyword demands and vague demands |
usingReRank | Enabled | Perform secondary sorting on initial recall results, prioritize displaying proposal segments with higher matching degrees with current customer demands, and improve the targeting of pre-sales Q&A |
This configuration is based on FastGPT v4.17.0
Three Common Misconfigurations
- Phenomenon: The locally deployed rerank model is not displayed in the interface. Cause: The access address and authentication information of the rerank model were not correctly configured on the FastGPT model management page, causing the system to fail to recognize the locally deployed service.
- Phenomenon: The reranking result is empty in the conversation details. Cause: The number of initial recall results does not meet the minimum input requirement of the rerank model, or the access timeout of the rerank model was not correctly configured.
- Phenomenon: The matching degree of retrieval results is low. Cause: The
similaritythreshold is set too high, filtering out some proposal content that meets the customer's potential demands, or hybrid retrieval mode is not enabled, relying only on a single retrieval method.
How to Verify Successful Configuration
- Navigate to the FastGPT model management page, and confirm that the rerank model configuration information matches the locally deployed service address and authentication parameters.
- Import a historical pre-sales technical solution document, perform hybrid retrieval and reranking operations, and check whether the returned result sorting matches the current test demand keywords.
- View the system logs to confirm that there are no timeout errors in the response time of the retrieval and reranking links.
- Adjust the
similaritythreshold, and verify whether the number and matching degree of recall results change as expected with the threshold adjustment.
The values provided 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.