What Data for This Category Looks Like
Data for this category comes primarily from design history files, component and material batch ledgers, manufacturing execution systems (MES), cleanroom environment monitoring equipment, sterilization equipment logs, and inspection and calibration systems. Daily documents include production process specifications, batch production records, cleanroom operation logs, sterilization batch records, inspection reports, software version records, adverse event and traceability materials. Data is updated per product batch or corresponding equipment unit. Cleanroom environment data is synchronized regularly per monitoring plans. Document fields include unique identifiers for critical materials, sterilization parameters, cleanroom temperature and humidity, software version numbers, and some records require association with batch and production step information.
Constraints on Knowledge Base Retrieval and Recall
Data for this category is updated per batch or monitoring plan. Retrieval must match the association between product batches and production steps to prevent incorrect data calls across batches. Multi-source, heterogeneous document formats require the retrieval workflow to support mixed recall of structured and unstructured data. Because this data relates to patient safety and product compliance, retrieval results must cover key control points, and recalled content must be traceable to original records. At the same time, the volume of content recalled in a single query must be kept within a reasonable length for compliant documents, to avoid exceeding information disclosure limits set by regulatory requirements.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Filters low-relevance non-control records and matches the relevance requirements of compliant documents for this category |
limit | Top 6–8 entries | Covers multi-node information for critical processes, sterilization, and cleanroom control |
usingReRank | Enabled | Sorts multi-source heterogeneous recall results by relevance, prioritizing display of key control parameters |
searchMode | mixedRecall | Supports joint retrieval of structured fields and text content, adapting to multiple document types |
datasetSearchUsingExtensionQuery | Enabled | Converts colloquial equipment inspection and maintenance questions into standardized queries that match production documents |
collectionIds | Bind dedicated collections for cleanroom production and equipment maintenance | Isolates documents exclusive to this category to avoid interference from irrelevant data in retrieval |
Three Common Configuration Mistakes
- Issue: After restoring a project backup, the agent configuration in the backend knowledge base is lost, and the specified retrieval collection cannot be called. Cause: The backup only restores project files and does not synchronize the associated
collectionIdsbinding configuration. The corresponding knowledge base collection must be rebound manually. - Issue: Knowledge base responses are truncated. Full content can be output normally in simple mode conversations. Cause: The number of recalled entries set by the
limitparameter is insufficient, or the retrieval results do not adapt to the context length of long documents, resulting in truncated model output. - Issue: After configuring a text understanding model, the re-ranking function does not work as expected. Cause: The configuration logic between the text understanding model and
usingReRankis confused. Re-ranking requires separately enabling the corresponding switch; the text understanding model is only used for the embedding generation stage.
How to Verify Successful Configuration
- Users log in to the FastGPT v4.17.0 backend, navigate to the corresponding knowledge base configuration page, and check that the set values of parameters such as
similarityandlimitmatch the configuration table. - Users enter a typical equipment inspection question, such as "What are the cleanroom temperature and humidity requirements for 10ml syringe assembly", and verify that the recall results include the corresponding batch's cleanroom records and production process specifications.
- Users trigger a single retrieval, check that the number of returned recall entries matches the
limitsetting, and that results are sorted by relevance. - Users verify that the
collectionIdsbinding configuration is retained after restoring a backup, to ensure that retrieval calls the dedicated knowledge base collection.
The values provided here are common starting points and should be evaluated against the reader's own samples.