What data looks like for this use case
Data for this use case comes from tender announcements, tender documents, enterprise qualification databases, and supplier quotations, updated per project. Daily processed documents include tender documents, qualification certificates, business deviation tables, and technical deviation tables. Most documents contain fixed-format clause response columns, qualification certificate fields, and quotation line items. These fields are tightly bound to the specific requirements of the tenderer, with no universal standardized format. All documents must strictly match the format template provided by the tenderer.
What constraints these characteristics impose on reranking and hybrid retrieval
Because data is updated per project and must match the tenderer’s exclusive format, retrieval must first recall content strongly relevant to the current tender clauses. This avoids mixing irrelevant qualifications or quotations into results. The fixed response columns in document structures require retrieval results to accurately correspond to clause numbers and requirements. Hybrid retrieval must balance text semantic matching and format keyword matching. The differences between the enterprise qualification database and supplier quotation data sources require the reranking step to distinguish credibility weights for different sources. This ensures qualification content prioritizes qualification review requirements, while quotation content matches business scoring items, preventing disordered sorting that disrupts subsequent document preparation.
How to set configurations
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Similarity threshold | 0.72–0.85 | Matches the tightly bound requirements of tender clauses, filters low-relevance non-tender content |
| Single search limit | Top 10 results | Controls the number of single-recall results, avoiding excessive irrelevant documents interfering with clause response verification |
| Retrieval mode | Hybrid retrieval | Combines semantic retrieval and keyword retrieval, covering both format keywords in tender announcements and semantic content of documents |
| Enable reranking | Yes | Reorders recall results by relevance to tender clauses, ensuring highly matched content is displayed first |
| Reranking recall count | Top 5 results | Focuses on highly matched content, adapting to the clause-by-clause verification needs of tender document preparation |
| Reranking model threshold | 0.75 | Distinguishes highly relevant and low-relevance recall results, ensuring reasonable sorting of qualification and quotation content |
Three common configuration mistakes
- Phenomenon: Reranking model fails to start, with a model loading timeout displayed on the interface. Cause: The local service port and model identifier are not correctly specified in the reranking model configuration, or the local model service is not running properly.
- Phenomenon: Reranked results do not follow business priority, with qualification content appearing after quotation content. Cause: The reranking weight is not adjusted for the qualification verification and business scoring needs of tender documents, and the default weight does not distinguish priorities between different data sources.
- Phenomenon: The number of retrieval results does not match the set single search limit, and the actual number of returned results is too low. Cause: The parameter linkage between hybrid retrieval and the reranking step is not correctly configured, causing some recall results to be incorrectly filtered during the reranking phase.
How to confirm configurations are correctly set
- Upload a test tender document and enterprise qualification database document, initiate a retrieval request for the corresponding tender clause, and check if the number of recall results matches the set single search limit.
- After enabling the reranking function, compare the result sorting before and after enabling, to confirm if qualification and quotation content follow business priority sorting.
- Check the reranking model operation logs, confirm there are no loading failures or timeout errors, and that the model normally returns reranking scores.
- Adjust the similarity threshold, verify that low-relevance content is correctly filtered, ensuring retrieval results only include documents strongly relevant to the current tender.
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.