What data in this category looks like
It typically serves data project managers, data engineers, and customer algorithm teams. Daily associated documents include dataset descriptions, version records, delivery checklists, and acceptance reports. Data sources cover object storage, data warehouses, annotation platforms, and customer acceptance feedback. Updates are made per batch delivery or version release. Data fields include version number, sample count, delivery path, difference description, acceptance status, and others. There are no unified fixed units; field rules must match the corresponding delivery package. The core focuses are the integrity of the delivery package and version traceability, which differs from single-sample validation in data cleaning and quality inspection stages.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Since data sources are scattered and linked to multiple types of business documents, retrieval must support unified recall across source indexes to avoid missing delivery data from different storage paths. The versioned update rhythm requires that retrieval results must be bound to the corresponding version identifier to prevent cross-version data mixing that causes training non-reproducibility issues. The fields of delivery checklists and acceptance reports have no fixed format, so retrieval configuration must support custom field mapping to ensure accurate matching of key information in delivery documents. At the same time, retrieval results must be traceable back to the original delivery package to meet customer verification requirements for data sources.
How to set the configuration
This configuration is based on FastGPT v4.17.0, and adapts to retrieval requirements for dataset delivery and version management scenarios.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Adapts to semantic matching accuracy for long texts such as dataset descriptions and version records, and avoids recalling irrelevant delivery data |
limit | Top 8 entries | Matches the document count range of a single batch delivery package, balancing retrieval efficiency and result coverage |
searchMode | vector | Prioritizes semantic recall, adapting to retrieval requirements for unstructured texts such as dataset version difference descriptions |
usingReRank | Enabled | Performs secondary sorting on recall results to improve the display priority of key information in pre-sales Q&A scenarios |
datasetSearchUsingExtensionQuery | Enabled | Combines the user's pre-sales consultation context to expand query terms, enhancing retrieval capabilities associated with context |
collectionIds | Specify specific IDs by delivery batch | Limits the retrieval scope to the current batch of datasets, preventing cross-version data mixing |
Three common mistakes
- Retrieval results include multiple versions of dataset samples. The cause is failure to use
collectionIdsto limit the dataset IDs of the specified version, leading to cross-version data mixing and causing training non-reproducibility or customer acceptance disputes. - Garbled characters or unrecognizable field content appear in retrieval results. The cause is failure to adapt to the original encoding format of datasets in object storage, and failure to specify corresponding encoding rules in the retrieval configuration, resulting in encoding exceptions during index construction.
- Unable to generate accurate retrieval results combined with the user's pre-sales consultation context. The cause is failure to enable the
datasetSearchUsingExtensionQueryparameter, and failure to expand queries using the user's context information, leading to retrieval results that do not match the current consultation scenario.
How to confirm the configuration is correct
- View retrieval logs to verify that the version information of returned results matches the specified
collectionIds, confirming that no irrelevant batch datasets are recalled. - Enter a query containing a version number to verify that the dataset description or delivery checklist document of the corresponding version can be accurately recalled.
- Test a pre-sales consultation query with context to confirm that the retrieval results combine context information and return delivery data related to the current consultation.
- Randomly select a retrieval result, check the content encoding format, and confirm that there are no garbled characters or missing field issues.
The values provided are common starting points and should be measured 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.