What the Data for This Category Looks Like
Typical entities include power transmission and distribution equipment, switchgear, and transformer manufacturers. Daily documents include winding and core assembly records, vacuum drying records, insulating oil reports, partial discharge and withstand voltage test reports, and nameplate parameter tables. Data sources include design BOM, silicon steel sheet and insulating material data, vacuum and drying equipment data, oil test results, and test station records, updated per product unit and material batch. Most documents follow a single-unit traceable structure, with fields such as insulation parameters, test values, and nameplate specifications. Units include megapascals, microamps, kilovolts, and other professional industrial parameter units.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
The single-unit traceable document structure requires retrieval to precisely match product batches and test items, to avoid generalized recall of irrelevant cross-batch data. Data updated per unit or batch requires regular refreshing of knowledge base shards, to prevent expired test records or material data from being included in retrieval results. Professional fields such as insulating oil and partial discharge tests have strict unit and numerical precision requirements, so retrieval must match both field names and numerical ranges. Otherwise, recall results cannot support ledger summarization. Multi-source heterogeneous data sources require unified field mapping, otherwise complete production process data cannot be associated, affecting the completeness of report generation.
Configuration Settings
Configuration settings are based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches the precision of insulation parameters and test values, to avoid recalling ledger documents from non-target batches |
limit | Top 6–10 entries | Adapts to the total number of documents for a single product, controlling the context length of retrieval results |
searchMode | mixedRecall | Balances retrieval effectiveness for structured BOM fields and semi-structured test reports |
usingReRank | Enabled | Re-ranks recall results by core business field priority, prioritizing return of core ledger items such as partial discharge and insulating oil test data |
datasetSearchUsingExtensionQuery | Enabled, extends retrieval scope to within 1000 tokens | Expands retrieval to associated equipment data and material parameters, improving information dimensions for ledger summarization |
collectionIds | Bind an enterprise-specific transformer ledger collection with a capacity not exceeding 1000 MB | Isolates documents from different production lines and batches, preventing cross-batch data confusion |
Three Common Misconfigurations
- Phenomenon: text-embedding-ada-002 has been added to a third-party API platform and the FastGPT configuration file has been updated, but a call to the knowledge base still prompts that the embedding model is unavailable or no matching results exist. Reason: The
collectionIdscorresponding to the embedding model have not been bound in FastGPT, or no exclusive documents for this category have been uploaded to the collection. - Phenomenon: The retrieved document reference fields are correct, but the generated production ledger report does not include the recalled test records or BOM data. Reason:
usingReRankhas not been enabled, resulting in too low a relevance ranking of recall results, which are not included in the valid context window. - Phenomenon: After initiating a conversation via API, the returned results do not associate with insulating oil reports or withstand voltage test data in the knowledge base. Reason:
datasetSearchUsingExtensionQueryhas not been enabled, so the retrieval scope only covers basic documents and does not associate equipment and material related data.
How to Confirm the Configuration Is Correct
- Enter the FastGPT knowledge base management page, verify that the
collectionIdsof the corresponding collection match the parameters in the configuration file. - Initiate a test retrieval, enter a query statement containing product batch and test type, verify that the document type and fields of the recall results match the business documents of this category.
- Enable FastGPT retrieval logs, check whether the
similarityscore of the recall results falls within the configured range. - Call the API to initiate a conversation, check whether the returned results include reference identifiers or related business fields of knowledge base documents.
The values provided are common starting points and should be measured against the reader's own samples.