What data for this category looks like
This category of data originates from the regulatory and rule repository, higher-level law database, departmental policy document repository, and administrative normative document filing system. Update frequency is irregular. Daily processed documents include normative document drafts, drafting explanatory notes, legality review opinions, solicited comment summaries, and filing reports. Data fields focus on policy-making authority, text wording, and procedural compliance. There are no unified fixed units. Match review basis based on the specific content of each document.
Constraints Imposed on Knowledge Base Retrieval and Recall by These Characteristics
Data sources are scattered, and updates follow no fixed cycle. Retrieval must cover all compliant basis repositories to avoid missing higher-level laws or latest policies. Document types vary, including long-text drafts and structured summary tables. Retrieval logic must adapt to different formats. The core of review focuses on policy-making authority and text legality. Retrieval results must accurately match review dimensions. A single review requires comparing multiple basis documents. The number and relevance of recall results directly affect review efficiency and accuracy.
Reference only; verify against the target FastGPT release and deployment documentation.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.88 | Matches precise compliant basis required for normative document review, avoids low-relevance results interfering with review judgments |
limit | Top 8–12 results | A single review requires comparing multiple basis documents. 8-12 results cover major higher-level laws and policy documents, avoids excessive results exceeding processing scope |
searchMode | Semantic retrieval | Adapts to review needs for abstract policy rules, accurately matches semantic content related to text legality |
usingReRank | Enabled | Performs secondary ranking on initial recall results, raises priority of results strongly related to review dimensions |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands query terms related to review, covers multi-dimensional retrieval scenarios such as policy-making authority and procedural compliance |
collectionIds | Bind dedicated compliant basis repository | Limits retrieval scope to repositories such as regulatory and rule repository and higher-level law database, avoids interference from irrelevant data |
Three Common Misconfigurations
- Retrieval time exceeds 4 seconds, and the interface displays a timeout status. Cause: Retrieval collection scope is not limited, full scanning of all repository data causes excessive retrieval load.
- Recall results include non-compliant basis documents, with empty fields or content unrelated to review dimensions. Cause:
collectionIdsconfiguration is incorrect, no dedicated compliant basis repository is bound. - Key content is missing from retrieval results for multi-row table documents. Cause: No adaptive processing for XLSX format documents, only plain text is extracted, leading to loss of table structure information.
How to Confirm Configurations Are Set Correctly
- Knowledge base retrieval logs are checked to confirm retrieval scope only includes preset compliant basis repository collections.
- A simulated review retrieval is initiated, and recall result similarity scores are verified to fall within the preset range.
- An XLSX format solicited comment summary document is imported, and comment content in the table is confirmed to be fully extracted during retrieval.
- Retrieval time is tested, and
limitandsimilarityparameters are adjusted to meet efficiency and accuracy requirements for business review.
The values provided are common starting points and 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.