What data for this category looks like
This specific category’s data comes from the centralized treasury payment system, unified budget management system, electronic invoice platform, and contract management system. Updates run daily. Daily processed documents include payment applications, fund appropriation plans, invoices, contracts, acceptance receipts, reimbursement approval slips, and payment vouchers. Document structures fall into three categories: approval process records, fund vouchers, and business contracts. Fields include business number, budget indicator code, economic classification code, amount, approval node status, and others. Single document lengths range from tens of characters to thousands of characters. This content uses the knowledge base feature of FastGPT v4.17.0.
What constraints these characteristics impose on knowledge base retrieval and recall
Accessing multiple data sources requires the retrieval pipeline to support unified indexing across data sources. Daily updates require configuring a scheduled synchronization mechanism to keep policy and business voucher data timely. Differences in document structure require chunking strategies adapted to semantic units of different document types, such as approval slips, contracts, and invoices. Single payment review needs precise matching of corresponding policy clauses to avoid irrelevant content interference. Thus, the relevance threshold for retrieval recall must be strictly controlled. All associated business data sources’ policy basis must be covered to prevent missed compliance judgments.
How to set configurations
| Configuration Item | Recommended Value | Basis for This Setting |
|---|---|---|
similarity | 0.75–0.85 | Requires precise matching of policy clauses and business scenarios, avoiding low relevance results that interfere with review judgments |
limit | Top 10 | Treasury review requires covering multi-dimensional policy basis, while controlling single-round context length to ensure response speed |
searchMode | vector + keyword hybrid | Policy text requires both precise keyword matching and semantic association needs; hybrid search balances recall completeness and relevance |
usingReRank | Enabled | Documents retrieved from multiple sources need further ranking calibration to ensure core policy clauses are returned first |
datasetSearchUsingExtensionQuery | Enabled | Can supplement omitted policy qualifiers in user questions to expand effective retrieval scope |
collectionIds | Exclusive dataset ID for treasury payment and reimbursement review | Isolate knowledge base data across different business scenarios to avoid irrelevant documents interfering with review processes |
Three common configuration errors
- Phenomenon: Single search takes more than 30 seconds, with delayed return results. Cause:
usingReRankis not configured, or thesimilaritythreshold is set too low, resulting in too many low relevance documents being recalled for reranking and sorting. - Phenomenon: Retrieval results only return content from a single knowledge base, and cannot display policy basis from two associated knowledge bases at the same time. Cause:
collectionIdsis not configured correctly, only a single dataset ID is specified, and all associated business data sources are not covered. - Phenomenon: Documents inserted via API cannot be retrieved in the knowledge base. Cause: The target
collectionIdsparameter is not specified, or the insert operation does not trigger dataset synchronization updates.
How to confirm the configuration is properly set
- Initiate a test query with a specific policy scenario, check whether the relevance of returned results meets review requirements.
- Call the API to perform a single retrieval, record the time taken and confirm it meets business response expectations.
- Insert a test document into the dataset with the specified
collectionIds, verify that the test content is included in the retrieval results. - Compare result sorting when
usingReRankis enabled versus disabled, confirm that the reranking logic takes effect.
The values given 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.