What this category of data looks like
Data for policy and regulatory basis Q&A for state-owned capital investment companies primarily comes from regulatory guidelines issued by the State-owned Assets Supervision and Administration Commission (SASAC), internal management systems of the group and its subsidiary companies, annual performance assessment documents, and compliance filing materials. Data updates trigger synchronously with the release of regulatory policies or revision of internal systems. There is no fixed update cycle, but core documents must be stored before they take effect. Each single document typically includes fields such as document number, issuing entity, effective date, scope of application, clause details, and attachment list. Document length varies widely, from 500 to 20,000 characters. Most fields are text-based, and some include structured performance indicator values.
What constraints do these characteristics impose on the knowledge base retrieval and recall workflow?
Scattered data sources covering SASAC regulatory documents, group internal systems, and subsidiary compliance documents require the retrieval system to split independent knowledge base collections by document hierarchy and type. Retrieval must limit collection scope to filter out irrelevant results. The lack of a fixed update cycle requires configuring support for incremental synchronization and manual trigger update mechanisms, to ensure the latest policies are stored in a timely manner. Documents include structured fields such as document number and effective date. Attribute filtering rules must be implemented to prioritize recalling policies that are within their validity period and match the applicable scope. Wide variation in document length requires balancing clause integrity during chunking, to avoid splitting core regulatory requirements and implementation details. This configuration was validated using FastGPT v4.17.0.
How to set the configuration
| Configuration Item | Recommended Value | Rationale for This Setting |
|---|---|---|
similarity | 0.72–0.85 | Policy documents require high matching accuracy, to avoid recalling non-core systems irrelevant to the business scenario |
limit | Top 10–15 results | Policy Q&A needs to cover relevant basis at different levels; too few results will miss key regulatory requirements |
searchMode | vector+keyword | Policy documents contain structured information such as clear document numbers and keywords; combining semantic retrieval improves recall accuracy |
usingReRank | Enabled | Semantic dispersion may occur after chunking long documents; re-ranking can place retrieval results of core clauses first |
datasetSearchUsingExtensionQuery | Enabled | Synonym associations exist in policy expressions, such as extended matching requirements between "state-owned capital operation" and "state-owned capital investment" |
collectionIds | Specify collection IDs for group-level and SASAC regulatory documents | Filter non-general policies of subsidiaries, narrow retrieval scope to improve result relevance |
Three common mistakes
- After calling the API to add or delete data, retrieval results are not updated synchronously. Cause: The
collectionIdsparameter is not specified correctly, resulting in data being written to a knowledge base collection not associated with the current retrieval task. - After passing in custom reference data, retrieval results are not linked to the corresponding original policy document. Cause: Reference fields are not passed in the structured format required by the system, or field names do not match preset reference matching rules.
- The first policy Q&A retrieval fails to recall the target file, but the second retrieval returns results normally. Cause: The vector index of the knowledge base has not been updated in real time; the first retrieval uses the old index cache, and the second retrieval triggers index refresh to recall correct results.
How to confirm the configuration is set correctly
- Upload a test group internal policy document, enter a query containing the core document number of the file, and check whether the retrieved results include this file.
- Adjust the value range of
similarity, observe changes in the matching accuracy of retrieval results, and confirm that the configuration item takes effect. - Switch between different modes of
searchMode, compare the coverage range of retrieval results, and verify the actual effect of the configuration. - Call the FastGPT dataset management API to add a test data entry, retrieve the corresponding keyword, and confirm that the data can be recalled normally.
Note: The configuration values provided are common starting points. Adjustments should be made based on the sample data and use cases of the respective deployment.
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.