What the data in this category looks like
Knowledge base data for state-owned capital operation companies mainly comes from group internal regulation compilations, normative documents issued by state-owned assets supervision departments, subsidiary operation manuals, employee compensation and benefits guides, office process SOPs, and similar materials. For update rhythm, state-owned assets policy documents update dynamically per regulatory requirements. Internal core regulations are revised as needed. Routine office documents are reviewed and updated in a centralized manner every six months. Most documents use official document formats. They include fields such as document number, issuing department, effective date, and scope of application. Process documents also include structured content like step nodes and responsible positions.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The authority of state-owned assets policy documents requires retrieval results to match the latest effective versions. Precise filtering must be applied to the document’s effective date field. This avoids recalling expired content. Internal regulations have clear scope of application. Retrieval must combine the user’s department and subsidiary to limit scope. This prevents returning documents that cross applicable scenarios. Official documents are mostly long texts. When chunking, key metadata such as document number and effective date must be retained. This avoids damaging the authoritative identification of the document. There are many structured fields. Relying only on keyword fuzzy matching cannot meet precise filtering needs. Targeted filtering by field must be supported.
How to set the configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value | Basis for This Value |
|---|---|---|
similarity | 0.72–0.85 | Documents for state-owned capital operation companies are mostly official texts. Semantic relevance requirements are high. This avoids low-match irrelevant content mixing into the result set. |
limit | Top 8–12 | Internal office documents are managed by business line and department. Too many results increase employee screening costs. Too few results may miss applicable content. |
searchMode | vector + keyword | Balances semantic matching and precise keyword hits. Adapts to retrieval needs for fixed fields like official document numbers and policy names. |
usingReRank | Enabled | Long text chunks may have scattered semantics. Reranking improves the accuracy of result relevance sorting. |
datasetSearchUsingExtensionQuery | Enabled | Expands synonymous expressions of user questions. This covers different terms for state-owned assets policies and internal regulations. For example, "compensation" and "salary and benefits". |
collectionIds | List of collection IDs divided by business line and department | Limits retrieval scope. This prevents recalling inapplicable documents across departments and subsidiaries. |
Three common errors
- Phenomenon: Retrieval results only return 1 text block. Cause: Incorrectly set the
limitparameter to 1. Did not adjust reasonable values based on business scenarios. - Phenomenon: Authorization failure error occurs when calling the knowledge base via the enterprise WeChat channel. Cause: Did not configure permission binding rules for the corresponding channel. Did not update adapted channel configuration parameters.
- Phenomenon: Retrieval results include expired state-owned assets policy documents. Cause: Did not enable filtering logic based on effective date fields. Did not limit retrieval to only active documents.
How to confirm the configuration is properly set
- Initiate a query containing a specific state-owned assets policy document number. Check if the retrieval results include the document with the corresponding number and rank it highly.
- Switch test accounts from different departments to initiate queries. Check if results only include documents applicable to that department.
- Adjust the value range of the
similarityparameter. Observe changes in retrieval result matching accuracy. Confirm the parameter takes effect. - Upload an expired test document and initiate a corresponding query. Confirm the document is not recalled.
The values provided are common starting points. They should be evaluated 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.