What the data for this category looks like
This category of data primarily comes from customer requirements, public data sources, collection tools, and internal sample libraries. Four types of documents are regularly produced: collection plans, sample specifications, authorization statements, and task schedules. These documents are updated per project or weekly. Data fields include sample scope, collection fields, compliance identifiers, and schedule nodes. Common units are sample counts, time cycles, and compliance status identifier fields. The core focus is defining rules and boundaries for raw data acquisition.
What constraints these characteristics impose on knowledge base retrieval and recall
Data for this category is updated per project or weekly. Retrieval and recall must support filtering by project and enterprise dimensions. This prevents cross-project or cross-enterprise data mix-ups. Documents include compliance identifiers. Retrieval must prioritize recalling collection rules with valid authorization. Most fields use specialized terminology such as sample scope and collection requirements. Retrieval must use domain-specific semantic matching. The sample scope definition process is time-consuming. Retrieval must quickly locate collection rule documents for the corresponding project. This reduces manual verification costs.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches semantic similarity for specialized documents such as collection plans and sample specifications, and avoids recalling irrelevant general technical documents |
limit | Top 3–5 entries | The number of collection rule documents per project is limited. Restricting the number of entries reduces context redundancy |
usingReRank | Enabled | For collection field descriptions with many specialized terms, re-ranking improves the sorting priority of highly relevant results |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonyms related to collection tasks. For example, expand "sample size" to "collected sample quantity" to cover more phrasing |
collectionIds | Fill in the knowledge base ID bound to the current project | Configuration must be completed under FastGPT v4.17.0. Restrict retrieval scope to only the current project's collection rule base to avoid cross-project data interference |
Three common mistakes
- Symptom: Image content in the knowledge base does not appear in retrieval results. Cause: Parameters supporting multimedia content retrieval are not configured, and only text snippets are recalled.
- Symptom: The knowledge base remains in the indexing state for a long time after upload. Cause: Retrieval scope is not limited by project or enterprise dimensions. The system scans too many non-target knowledge base datasets, causing indexing progress blockage.
- Symptom: Results with high full-text retrieval scores rank lower than results with lower scores. Cause: Re-ranking function is not enabled. Only basic similarity sorting is used, and no secondary calibration is performed for specialized semantics in the collection domain.
How to confirm correct configuration
- Initiate a retrieval targeting the current project's collection plan. Verify that returned results only include documents for the current project, with no cross-project or irrelevant content.
- Upload a collection rule document with images, then initiate a retrieval. Verify that image content is included in the results.
- Adjust the similarity threshold, then initiate a retrieval. Observe whether the number of returned results changes as expected with the threshold adjustment.
- Enable the re-ranking function, then initiate a retrieval. Verify that results with high semantic relevance appear at the top of the list.
The values provided here 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 2026-10-19.