Reference only; verify against the target FastGPT release and deployment documentation.
What This Category of Document Data Looks Like
Document data for state-owned asset management companies mainly comes from policy documents issued by state-owned asset supervision departments, internal asset transfer agreements, equity disposal plans, annual state-owned asset operation reports, risk control and compliance review documents, and similar materials. Update frequency adjusts based on regulatory requirements and project progress, with no fixed cycle. Most documents include uniform document numbers, asset subject information, amount fields, and approval process nodes. Common field units are ten thousand yuan, hundred million yuan, number of households, and similar units.
What Constraints Do These Characteristics Impose on the Knowledge Base Retrieval and Recall Link
Individual documents have long length, which easily causes complete clauses to be truncated during chunking. There are many structured fields, so retrieval must match both semantics and specific fields such as document numbers and amounts. Pure semantic recall may miss content that matches exactly. Updates have no fixed cycle, so incremental synchronization mechanisms must be supported. The sensitive nature of documents requires strict restriction of the retrieval set to avoid unrelated documents appearing in search results.
How to Set the Configuration
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | State-owned asset management documents are mostly rigorous clause-based content. This range filters low-match irrelevant results while retaining matching content for core clauses |
limit | Top 8 results | Individual documents have long content. Too many retrieved results will exceed the context window, while this number covers different clause fragments |
searchMode | semantic+keyword hybrid mode | Balances semantic matching and precise matching for structured fields such as document numbers and amounts |
usingReRank | Enabled | Semantics may be scattered after long documents are chunked. Reranking filters the most relevant fragments |
datasetSearchUsingExtensionQuery | Disabled | Document drafting requires strict matching of original clauses in the knowledge base. Avoiding AI extension prevents deviation from preset content |
collectionIds | Specify collection IDs for corresponding state-owned asset management documents | Restrict retrieval scope to avoid interference from unrelated documents on search results |
Three Common Mistakes
- Phenomenon: Text chunks are lost when chunk length is set to 3000 characters. Cause: Most state-owned asset management documents contain continuous clause paragraphs. A 3000-character chunk may cross complete clause boundaries, leading to truncation of core clauses and loss of complete information.
- Phenomenon: An error indicating rate limit exceeded occurs during vectorization. Cause:
collectionIdsis not specified, leading to batch processing of a large number of unrelated documents and exceeding the concurrency limit of the embedding interface. - Phenomenon: Search results deviate from the original knowledge base content, with additional extended content. Cause: The
datasetSearchUsingExtensionQueryparameter is enabled, causing the AI to perform unnecessary semantic expansion on retrieved content.
How to Confirm the Configuration Is Correct
- Upload a complete internal asset disposal plan document, check the chunk preview, and confirm that no complete approval clauses or amount fields are truncated.
- Initiate a retrieval targeting a specific state-owned asset supervision document number, confirm that only documents within the specified collection are retrieved, with no unrelated results across collections.
- Test drafting a document based on retrieval results, confirm that all cited content comes from the original knowledge base, with no expressions not present in the knowledge base.
- Check the embedding task logs, confirm that the number of concurrent requests does not exceed the interface limit, and there are no rate limit exceeded errors.
The values provided here are common starting points. They should be evaluated against your own specific samples and requirements.
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.