What data for this category looks like
This category’s data draws from design BOMs, silicon steel sheet and insulating material ledgers, vacuum drying equipment logs, oil test reports, partial discharge and withstand voltage test records from test stations, and product nameplate parameter tables. Daily documents update per single unit or batch. They include winding and core assembly process parameters, vacuum degree and holding duration, insulating oil moisture content and breakdown voltage, partial discharge peak values and withstand voltage values, and tap changer positions. Units include professional metering values such as Pascals, degrees Celsius, megavolts, and picocoulombs. Full-process data must be associated using the unique product identifier.
Constraints imposed on retrieval and recall workflows
The update frequency of this category’s data at the single unit or batch level requires retrieval and recall to filter the recall scope by the unique product identifier. This prevents mixing data across different products. Professional metering fields and power industry-specific terms such as partial discharge peak values and insulating oil breakdown voltage require the retrieval model to adapt to industry semantics. This stops generic retrieval from matching irrelevant documents. Multiple heterogeneous data sources require retrieval to support cross-collection recall. It also requires strict control of the number of recall results, to avoid excessive non-core data interfering with inspection and review processes.
How to set configurations
These configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches this category’s professional semantic retrieval needs, balances recall precision and scope. Prevents too many irrelevant results from low thresholds, and filters out qualified relevant data from overly high thresholds |
limit | Top 8–12 results | Controls the number of results returned per retrieval round, adapts to the information density of inspection and review scenarios, and avoids overly long contexts exceeding model processing limits |
searchMode | embedding | Adapts to matching needs for this category’s professional terms and unstructured documents, and improves cross-document semantic association capabilities |
usingReRank | Enabled | Performs secondary ranking on initial recall results, corrects semantic deviations from initial retrieval, especially for matches of professional indicators such as partial discharge and insulating oil |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval queries, supplements synonyms related to power equipment industry terms, such as expanding shorthand references to the standard term, and improves professional term matching rates |
collectionIds | Configured using document collection IDs associated with the product | Limits the retrieval scope to only document collections for the current batch or single product, avoiding cross-product data interfering with reviews |
Three common configuration mistakes
- Phenomenon: When the
similaritythreshold is set to 0.77, the system prompts that no relevant information is found. Cause: Semantic vector matching for this category’s professional terms has relatively high difficulty. An overly high threshold filters out qualified documents that are semantically related but have slightly lower similarity. - Phenomenon: Retrieval results only cover a single type of document, and cannot recall both insulating oil reports and partial discharge test records at the same time. Cause: The multi-
collectionIdsparameter is not configured, and the retrieval scope is limited to a single document collection only. - Phenomenon: The AI output content includes rewritten statements not taken from the original source, and does not directly output the original knowledge base text. Cause: The native content direct return logic for retrieval results is not enabled. The system automatically performs semantic reconstruction on the recalled content.
How to confirm the configuration is correct
- Enter inspection questions that include professional terms, such as "Does the insulating oil moisture content of this batch of transformers meet the requirements?" Confirm that retrieval results cover the corresponding type of documents.
- Count the number of retrieval return results, and confirm that they fall within the range set by the
limitparameter. - Verify that multiple sources of process records can be recalled at the same time, and confirm that the multi-
collectionIdsconfiguration takes effect. - Confirm that AI output content directly references original knowledge base text, without adding additional descriptive statements.
The values provided are common starting points and should be measured against the reader's own samples.