What This Category's Data Looks Like
The data used for document and material drafting at state-owned capital investment companies primarily comes from internal compliance manuals, state-owned assets supervision policy documents, investment project feasibility study reports, annual operation summaries, equity structure documents, and similar materials. The update rhythm adjusts based on state-owned assets supervision requirements and project progress. Policy documents are updated quarterly, while project documents are updated dynamically alongside project approval, review, and implementation milestones. Most documents follow formal official document formats, including structured fields such as document numbers, project IDs, supervision levels, and some contain numerical content such as investment amounts and shareholding ratios. Document length varies widely, ranging from a few pages of compliance guidelines to dozens of pages of project reports.
Constraints Imposed on Knowledge Base Retrieval and Recall
First, the wide range of document lengths — from a few pages of compliance guidelines to dozens of pages of project reports — creates significant differences in single-block text length. This requires a flexible chunking strategy to adapt to different document types. Second, documents include structured metadata such as document numbers, project IDs, and supervision levels, as well as numerical content such as investment amounts and shareholding ratios. This demands support for metadata filtering and numerical range retrieval to avoid retrieving irrelevant content. Third, policy documents require strict alignment with regulatory requirements. Retrieved results must cover the latest released regulatory documents. Meanwhile, project documents need to be linked to full-cycle materials for corresponding projects. This places requirements on the knowledge base’s metadata association and update synchronization capabilities.
The following configurations apply to FastGPT v4.17.0.
How to Configure the Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Document drafting for state-owned capital investment companies requires strict alignment with regulatory policies and internal specifications. A threshold that is too low will introduce irrelevant content, while a threshold that is too high may miss relevant compliance clauses. Adjust based on the compliance scenario interval calibrated through actual testing. |
limit | Top 8–12 results | The total number of retrieved results for project documents and policy documents must cover the core references required for drafting. Too many results will exceed the context window limit, while too few will fail to provide sufficient reference materials. |
searchMode | Hybrid search | Users need both semantic matching for policy text and precise matching for structured metadata such as document numbers and project IDs. Hybrid search balances semantic relevance and precise matching needs. |
usingReRank | Enabled | Document drafting has high requirements for content relevance. Reranking filters out results with weak semantic association from initial retrieval, improving the quality of final retrieved content. |
datasetSearchUsingExtensionQuery | Enabled | Document drafting for state-owned capital investment companies often requires linking multiple relevant policy and project documents. Extended queries can supplement implicitly associated content for drafting scenarios, covering more necessary reference materials. |
collectionIds | Bind knowledge base collections corresponding to state-owned assets supervision, project management, and compliance documents | Accurately limit the retrieval scope, avoid retrieving irrelevant non-business documents, and improve retrieval efficiency and relevance. |
Three Common Misconfigurations
- Issue: An unsupported return format error occurs when calling the knowledge base retrieval interface, but the original document link can be downloaded normally as a docx file in a browser. Cause: The document parsing whitelist is not configured, or the uploaded docx file contains encrypted content or nested macros, leading to parsing failure.
- Issue: The number of retrieved results does not match the configured
limitparameter, with either too many or too few returned results. Cause: ThecollectionIdsparameter is not correctly bound, so the retrieval scope covers non-target knowledge base collections, or the return upper limit for extended queries is not restricted after enablingdatasetSearchUsingExtensionQuery. - Issue: The function that allows users to choose whether to skip knowledge base retrieval cannot be implemented. Cause: Dynamic variable trigger logic is not configured, and the retrieval switch is not bound to the interface call condition, making it impossible to adjust the retrieval process based on user selection.
How to Verify Successful Configuration
- Upload a test state-owned assets supervision policy document and a project feasibility study report, initiate a relevant retrieval request, and verify whether the retrieved results include the core content of the target documents.
- Check the retrieval logs to confirm that the actual values of configuration parameters such as
similarityandlimitmatch the preset values, with no missing parameters or incorrect bindings. - Test enabling and disabling
usingReRank, compare the relevance differences between the two retrieval results, and verify that the reranking function works as expected. - Attempt to trigger the scenario where retrieval is skipped, confirm that the knowledge base retrieval request is not initiated when the interface is called, and directly execute the preset process content.
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.