What the data for this category looks like
Data mainly comes from internal service process manuals of state-owned capital investment companies, policy documents released by state-owned assets supervision departments, archived past service consultation tickets, and service guidelines submitted by subordinate enterprises. Update timing triggers on demand when state-owned assets policies are adjusted or internal processes are optimized, with no fixed cycle. Most individual documents are structured content, including fields such as service item name, handling subject, declaration material list, handling time limit, consultation channel, and approval node. Handling time limits are measured in working days. Declaration material lists are counted in units of "items".
Constraints imposed by these characteristics on the knowledge base retrieval and recall link
The structured document characteristics of this category require retaining complete semantics within fields during retrieval. Avoid splitting associated information across service items or declaration materials, to prevent context breaks in recall results. The mixed attribute of multi-source data requires distinguishing the authoritative weight of state-owned assets policy documents from the practical reference weight of past consultation tickets. This prevents practical content from overriding policy requirements. The feature of no fixed update cycle requires the knowledge base to support incremental synchronization mechanisms. Updated latest policies and process adjustments can be made without full reconstruction. Documents involving differentiated service rules of subordinate enterprises need to filter recall results by handling subject dimension, to ensure results match corresponding business subjects.
Configuration setup
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration item | Recommended value | Basis for this setting |
|---|---|---|
similarity | 0.72–0.80 | Service consultation requires accurate matching of business rules. Too low a threshold will introduce irrelevant service content, too high will miss compliant results, calibrated based on actual testing. |
limit | Top 8 entries | Valid results for service consultation are usually concentrated in a small number of items. Too many results will exceed the context window and affect the accuracy of the large model's responses. |
searchMode | vector | Adapts to the field matching requirements of structured documents. Vector retrieval can accurately match keywords and semantically related content, ensuring the recall accuracy of service items. |
usingReRank | Enabled | Results for service consultation need to be sorted by relevance. Reranking can correct local deviations from vector retrieval, ensuring core fields such as declaration materials and handling time limits appear first in results. |
datasetSearchUsingExtensionQuery | Enabled | User questions in past consultation tickets have synonymous expressions. Extended queries can cover more actual user question scenarios and improve recall coverage. |
collectionIds | Bind dedicated service consultation dataset ID | It is necessary to isolate service consultation-related documents from other business datasets, avoid recalling irrelevant content, and ensure the accuracy of the retrieval scope. |
Three common configuration mistakes
- Phenomenon: A connection timeout error occurs when binding the service consultation dataset after deploying FastGPT. The log shows that the dataset index construction failed. Cause: No reasonable chunking parameters are set for structured documents, resulting in single chunk content exceeding the processing limit of the vector model, causing index construction exceptions.
- Phenomenon: Retrieval results include service rules across subordinate enterprises, and the returned results do not match the handling subject in the user's question. Cause: The
collectionIdsconfiguration to limit the retrieval scope is not set, resulting in recall of content from other business datasets and confusing service rules of different subjects. - Phenomenon: The declaration material list in the retrieval results is incomplete, only showing some entries. Cause: The associated content of the declaration material list was split during chunking, destroying the complete semantics of structured fields, resulting in missing key information in the recall results.
How to confirm the configuration is correct
- Manually enter a typical service consultation question, verify that the handling subject of the returned results matches the question, and confirm that the
collectionIdsconfiguration takes effect. - Adjust the
similaritythreshold, observe changes in the relevance of retrieval results, and confirm that the threshold setting meets business requirements. - Enable
usingReRank, compare the result sorting before and after enabling, and confirm that the reranking function is operating normally. - Enter questions with synonymous expressions, verify that documents corresponding to the service items are recalled, and confirm that the
datasetSearchUsingExtensionQueryconfiguration takes effect.
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.