What this category’s data looks like
Typical data for this category comes from enterprise IT procurement systems, supplier quotation documents, budget spreadsheets, and contract systems. Daily documents include purchase requisitions, price comparison records, formal contracts, and acceptance reports. Purchase requisitions typically contain requirement specifications, application amounts, and user department information. Price comparison records include quotations and qualification documents from multiple suppliers. Contracts contain clause details and performance deadlines. Acceptance reports include delivery and performance status. Data update frequency follows procurement order execution. Single procurement-related documents update per order or monthly. Qualification documents for some long-term suppliers update quarterly.
What constraints do these characteristics impose on reranking and hybrid retrieval
This category’s data includes both structured fields and non-long-form content. Structured fields on purchase requisitions must be matched first. Long-form content from contracts and price comparison records requires precise keyword matching. Multi-source data has uneven update frequencies. Some historical contracts and qualification documents update slowly, so priority must be differentiated during retrieval. Internal work order questions mostly focus on procurement requirements, supplier qualifications, and contract clauses. Both structured procurement data and unstructured documents must be retrieved. The reranking link must balance relevance sorting for different data formats, to avoid low-relevance results interfering with work order responses. Procurement-related retrieval results directly affect procurement process progress. Reranked results must accurately match user questions.
How to set the configuration
This explanation is based on FastGPT v4.17.0.
| Configuration Item | Suggested Setting | Rationale |
|---|---|---|
| Rerank Toggle | Enabled | Procurement documents often include long-form clauses and structured fields. Reranking effectively filters low-relevance retrieval results |
| Similarity Threshold | 0.72–0.85 | Matches core fields for procurement requirements, avoids retrieving irrelevant historical contracts or supplier quotations |
| Single Retrieval Limit | Top 8–12 entries | Balances response speed and retrieval coverage for internal work order Q&A, avoids excessive redundant data increasing context length |
| Retrieval Mode | hybrid | Combines vector retrieval and keyword retrieval, supports hybrid retrieval for structured procurement system data and unstructured documents |
| Rerank Model Call Timeout | 60 seconds | Accommodates long-text processing for procurement documents, avoids work order Q&A timeouts due to slow model responses |
| Rerank Model Selection | Set based on actual testing | Adapts to the text characteristics of enterprise procurement documents. Prioritize rerank models that support long-text processing |
Three common mistakes
- Phenomenon: Unable to select the rerank model when creating an application. The interface shows blank or throws an error. Cause: The API address and key for the rerank model were not correctly configured in config.js, or the configuration format does not meet FastGPT requirements.
- Phenomenon: A large number of irrelevant supplier quotations or historical contracts appear in retrieval results, and the number of results does not match expectations. Cause: A reasonable similarity threshold was not set, or reranking to filter low-relevance results was not enabled, leading to redundant retrieval results.
- Phenomenon: The rerank model call returns
rerank error fastgpt | { message: { detail: ' error message. Cause: The offline rerank model was not deployed correctly, the API call timed out, the key is invalid, or the rerank model access configuration was not completed per documentation.
How to confirm the configuration is correct
- Navigate to the FastGPT application configuration page, confirm that the rerank toggle is enabled, and parameters such as similarity threshold and single retrieval limit are set to the corresponding values.
- Submit a test work order, enter a procurement-related question, and check whether the number and relevance of retrieval results meet expectations.
- View the application’s running logs, confirm that the rerank model call was successful with no error messages.
- Adjust the similarity threshold, verify that the relevance of retrieval results adjusts appropriately as the threshold changes.
All configuration values listed here are common starting points. Adjust them based on your own sample data and testing.
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.