What the data for this category looks like
Data for this category comes from the China Government Procurement Network, local government procurement websites, fiscal integration systems, purchaser budget systems, and supplier credit records. The update frequency is daily. Daily documents include procurement requirements, procurement plans, bidding documents, tender documents, bid evaluation reports, winning bid announcements, procurement contracts, and question responses, among others. Most fields include project numbers, procurement content, compliance clauses, supplier information, and evaluation results, among others. Document structure varies significantly based on business process differences.
What constraints these characteristics impose on the vector models and indexes link
Data is updated daily and sourced from multiple dispersed locations. Incremental indexing workflows for multi-source heterogeneous documents must be supported to avoid full repeated indexing that consumes system resources. There are diverse document types. Core fields of different documents (such as compliance clauses in procurement requirements, evaluation conclusions in bid evaluation reports) require targeted extraction. Chunking rules must match text length to preserve semantic integrity. For compliance content related to government procurement, indexes must retain the accuracy of key clauses to avoid incorrect recall of non-compliant content. Data volume increases gradually as procurement projects expand. Resource usage of indexing scheduling must be controlled to ensure normal operation of business processes.
How to set the configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the paragraph length of government procurement documents, preserves semantic integrity of key compliance clauses |
| Overlap Length | 100–150 characters | Prevents key clauses from being truncated after chunking, ensures semantic coherence across chunks |
maxToken | 8191 | Adapts to the maximum token limit of mainstream embedding models such as text-embedding-ada-002, avoids vector generation failures |
| Single Search Limit | Top 8 entries | Matches the range of associated document counts for government procurement projects, balances recall efficiency and coverage |
| Similarity Threshold | 0.72–0.78, calibrated via actual testing | Differentiates party building learning-related content from non-related procurement documents, reduces false recall probability |
defaultToken | 1536 dimensions | Matches the output dimension of mainstream models such as text-embedding-ada-002, ensures indexing compatibility |
Three common mistakes
- The interface displays "No matching vectors" or an error occurs during the indexing process. Cause: The embedding model is not correctly associated with the vector channel of the current knowledge base, or the configured
defaultTokendimension does not match the model's output dimension. - Indexing tasks remain in the "Indexing" status with no progress. Cause: The locally deployed embedding model has an incorrectly configured API access path, or the chunk length setting exceeds the model's
maxTokenlimit, resulting in generation timeout. - No valid results are returned after creating a data index. Cause: Chunking rules are not configured based on document type, leading to incorrect extraction of core fields from government procurement documents, making it impossible to generate valid vectors.
How to confirm the configuration is correct
- Upload a single fragment of a government procurement bidding document, check the vector generation log, and confirm that the vector dimension returned by the embedding model matches the configured value.
- Trigger an incremental indexing task, check the system task queue, and confirm that the task completes within a reasonable duration with no error logs.
- Input a test question related to party building learning, check the recall results, and confirm that all returned documents are related to government procurement content, with no unrelated documents included.
- Check the vector channel configuration of the knowledge base, and confirm that the added embedding model matches the currently used model, with no configuration omissions.
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 2026-09-28.