What data for this category includes
Data for unit-based system retrieval comes from official internal documents of state-owned enterprises and government departments, including departmental regulations, operating guidelines, official documents with red headers, and more. Update frequency is determined by each unit’s internal system release process. It is synchronized when a new system takes effect, with no fixed cycle.
Reference only; verify against the target FastGPT release and deployment documentation.
Constraints imposed by these characteristics on the knowledge base retrieval and recall link
Because documents are tied to a specific issuing authority, retrieval must strictly limit the unit scope to prevent accidental recall of cross-unit systems. Range filtering based on unit identifiers is required.
Slight differences in document fields require unified preprocessing of unit identifier fields before retrieval, to ensure consistent matching logic. The lack of a fixed update cycle requires index updates to support on-demand triggering. Fixed-cycle synchronization cannot adapt to newly released systems that are issued temporarily.
Wide variation in individual document lengths requires retaining unit identifiers during chunking, to avoid losing the owning unit information after chunking, which affects attribution judgment for subsequent recall results.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale for This Setting |
|---|---|---|
similarity | 0.75–0.85 | Unit-based system content is specialized and has clear boundaries. A higher threshold filters irrelevant documents with low matching scores |
limit | Top 8–12 results | Individual system documents are lengthy. Too many recall results will exceed the conversation context capacity limit |
searchMode | Vector search + keyword hybrid | Systems contain specialized terminology and clear document numbers and unit identifiers. The hybrid mode balances semantic matching and precise keyword matching |
usingReRank | Enabled | Long documents may have scattered semantics after chunking. Reranking corrects ranking biases from initial recall |
datasetSearchUsingExtensionQuery | Disabled | Internal system Q&A mostly involves precise business queries. Extended queries may introduce irrelevant business scenarios |
collectionIds | List of dataset IDs bound to the current unit | Core configuration for implementing unit-based retrieval. Limits retrieval scope to only system documents from the corresponding unit |
Three common configuration mistakes
- Symptom: Retrieval results include system documents from other units, and returned content is unrelated to the queried unit. Cause:
collectionIdsis not configured correctly, and the retrieval dataset scope is not limited, resulting in cross-unit documents being included in recall results. - Symptom: Retrieved content does not match the queried system clauses, but instead links to other irrelevant business content. Cause: The
similaritythreshold is set too low, and non-target documents with low matching scores are prioritized for recall. - Symptom: Retrieval results do not cover the full system main text, only returning scattered chunked content. Cause: The
limitvalue is too small, and not enough relevant chunks are recalled, resulting in key system content not being detected.
How to verify the configuration is correct
- Submit a dedicated system query for the target unit, and verify that retrieval results only include system documents from that unit, to confirm the scope filtering configuration works.
- Adjust the
similaritythreshold, observe changes in retrieval result matching scores, to confirm the threshold meets the precision requirements of the business scenario. - After enabling
usingReRank, compare the initial recall results with the reranked result order, to confirm the reranking function operates normally and optimizes the ranking logic. - Enter precise keywords such as system document numbers, to verify that the hybrid retrieval mode can quickly match target documents.
The values provided here 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 2026-09-28.