This document is compatible with the knowledge base retrieval and recall logic of FastGPT v4.17.0.
What this category of data looks like
Data for this category primarily comes from object storage, data warehouses, annotation platforms, and customer acceptance feedback. Updates are made per batch delivery or version release. Daily documents include dataset descriptions, version records, delivery lists, and acceptance reports. Core fields include version number, sample coverage, delivered file details, acceptance status, and update notes. There are no uniform mandatory units. Some fields must match business identifiers provided by customers. Data management requires strict version tracing to avoid mixing samples from different batches.
What constraints these characteristics impose on knowledge base retrieval and recall
Data for this category is updated in batches per delivery or version, and must be strictly traced via version numbers. Knowledge base retrieval must support filtering recall scope by version. Datasets from different versions have differing samples. Retrieval must avoid mixing data across versions. Metadata fields for documents such as delivery lists and acceptance reports have no fixed unified format. They must adapt to multiple types of business identifiers. Recall must support dynamic matching of metadata tags. Data updates occur at a high frequency. Indexes must be synchronized regularly to ensure retrieval results match the latest delivered versions, and prevent returning expired dataset content.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches structured metadata such as dataset version numbers and sample descriptions. This range filters low-relevance results and prevents cross-version data recall |
limit | Top 8–12 entries | The number of dataset documents delivered in a single batch is typically limited. This range covers retrieval results for complete delivery packages and avoids redundancy |
searchMode | vector+keyword | Balances semantic content of dataset descriptions with keyword matching for version numbers and delivery lists, improving retrieval accuracy |
usingReRank | enabled | Dataset metadata fields are numerous and structurally flexible. Reranking further filters high-match version records and delivery details |
datasetSearchUsingExtensionQuery | enabled | Supports supplementary matching of business identifier fields provided by customers, adapting to different customers' dataset delivery formats |
collectionIds | Specify knowledge base collections bound to the delivered version | Strictly limits retrieval scope, prevents recalling dataset data from other versions, and ensures accurate version tracing |
Three common mistakes
- Retrieval results return cross-version dataset content. For example, searching for "v1.0 delivery list" returns v2.0 documents. The cause is failing to bind the corresponding version's
collectionIdsconfiguration, and failing to limit the retrieval scope. - Insufficient retrieval accuracy, unable to match dataset descriptions or delivery details for a specified version. The cause is failing to enable
datasetSearchUsingExtensionQuery, and failing to supplement matching of customer business identifier fields, resulting in failure to correctly recall structured metadata. - The number of results returned by a single retrieval exceeds expectations. For example, an API call returns more than 20 dataset documents. The cause is failing to correctly set the
limitparameter, with a value range that is too large, leading to excessive redundant results.
How to confirm configurations are set correctly
- Enter a search term with a specified version number, check if returned results only include dataset documents for that version. Verify that the
collectionIdsandsimilarityconfigurations are effective. - Search for a compound term that includes a version number and delivery details, check if results recall both semantic matches and keyword matches. Verify that the
searchModeconfiguration is correct. - View the ranking of retrieval results, confirm that high-match version records and delivery lists appear at the top. Verify that the
usingReRankconfiguration is enabled. - Enter a search term for an unbound version number, check if no matching results are returned. Verify that the retrieval scope limitation is accurate.
The values provided here 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 2026-10-19.