What the data for this use case looks like
The data sources for this use case are tender notices, tender documents, enterprise qualification libraries, and supplier quotation libraries, with an update frequency of once per project. Daily processed documents include tender documents, qualification certificates, business deviation tables, and technical deviation tables. Data fields include qualification items required by the tenderer, quotation details, clause response descriptions, deviation reasons, and more. The overall structure strictly matches the format sections of the tender notice. The number of response items per document varies with the scale of the tender.
What constraints do these characteristics impose on reranking and hybrid retrieval?
This use case’s data is structured according to standard tender formats, and each document contains a large number of response items. This requires that reranking and hybrid retrieval must first prioritize matching qualification and clause requirements explicitly listed by the tenderer. Since data is updated once per project, the dataset for each retrieval is a temporarily generated project-specific library, so long-term persistent indexing is not needed. Additionally, data from different sources has significant format differences. Hybrid retrieval combining vector similarity and keyword matching results is required to cover different types of business fields such as qualification certificates and quotation details, and avoid missing key information due to a single retrieval mode.
How to configure the settings
The following configurations are compatible with FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Similarity Threshold | 0.75–0.85 | Filter redundant data unrelated to tender clauses and qualification requirements, avoid tender document response errors that could lead to bid rejection risks |
| Single Search Limit | Top 10 results | Cover multi-response item requirements for a single tender document, avoid missing key qualifications or clauses, and ensure all tender requirements are addressed |
| Search Mode | Hybrid Retrieval | Combine vector retrieval and keyword retrieval to adapt to multi-source data such as tender notices and qualification libraries, and ensure coverage of business information in different formats |
| Result Reranking | Enabled | Reorder initial retrieval results by business relevance, prioritize content matching tender requirements, and improve business processing efficiency |
| Reranking Timeout | 3–5 seconds | Adapt to the retrieval speed of temporary project-specific datasets, avoid request failures caused by overly long reranking delays that could impact tender document preparation progress |
Three common configuration errors
- Phenomenon: Query speed slows significantly after adding a reranking model, and the API returns a timeout prompt. Cause: The single search limit for initial retrieval was not adjusted. Excessive initial results cause the reranking model inference time to exceed the threshold, extending the total retrieval process duration.
- Phenomenon: After enabling result reranking for knowledge base search, the returned result order is identical to the initial vector retrieval results. Cause: The reranking model was not correctly bound to the current application, or the result reranking configuration switch was not enabled, causing the reranking logic to not take effect.
- Phenomenon: The reranking model deployed via Docker runs successfully, but the returned reranking results are false or empty. Cause: The input format passed to the reranking model does not meet requirements, and the query text and initial retrieval result list were not correctly transmitted.
How to confirm successful configuration
- Access the FastGPT application configuration interface, confirm that the result reranking switch is enabled.
- Initiate a test retrieval, compare the initial vector retrieval results with the reordered results to confirm that the results have been adjusted by business relevance.
- Check the running status of the reranking model, confirm that the Docker container is running normally and there are no API call errors.
- Verify the filtering effect of the similarity threshold, confirm that low-similarity redundant data has been excluded.
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.