What the data for this category looks like
This category’s data is sourced from cell banks and seed lots, process analysis, cleanroom monitoring, MES, LIMS, and equipment integrity testing. Update cadence aligns with each production batch, real-time monitoring data, and fixed-cycle planned execution. Daily documents include cell bank or seed lot records, culture and purification records, aseptic filling records, environmental monitoring data, media fill simulation reports, sterility test results, and batch release documentation. Data fields include batch number, culture temperature, cleanroom settleable plate count, equipment operating duration, and more, with corresponding units such as degrees Celsius, CFU/m³, hours, and others.
Constraints imposed on knowledge base retrieval and recall by these characteristics
The multi-source, high-frequency update nature of category data requires knowledge base retrieval to support filtering recall scope by production batch and monitoring cycle, to avoid calling expired process data. Documents contain large volumes of professional process parameters and compliance records. During retrieval, content related to aseptic control and batch release for bidding scenarios must be prioritized to prevent generalized recall of irrelevant data. Data update frequencies vary widely. Real-time monitoring data must be synchronized to the knowledge base, and the latest cleanroom environment and equipment integrity data must be prioritized during retrieval to ensure bidding content meets current compliance requirements. Due to the risk profile of aseptic manufacturing, retrieval results must have high confidence to avoid missing critical process parameters.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Similarity threshold | 0.75–0.85 | Match professional documents related to aseptic processes and batch release, avoid recalling irrelevant general production content |
| Maximum single search results | Top 8–12 entries | Cover the number of process parameters and compliance records required for a single bid document, while controlling retrieval latency |
| Search mode | Hybrid vector + keyword | Combine semantic matching and keyword retrieval to accurately recall documents containing specific fields such as batch number and settleable plate count |
| Reranking toggle | Enabled | Rearrange initial recall results by relevance to ensure the most relevant aseptic control data is prioritized |
| Query optimization toggle | Enabled | Automatically expand query terms for bidding-related process questions to cover different expressions of professional terminology |
| Designated knowledge base collection | Bound to this category’s exclusive knowledge base | Limit retrieval scope to only compliance and process documents for aseptic manufacturing, avoid interference from cross-category data |
Three common mistakes
- Phenomenon: Knowledge base search returns empty results, or only returns general documents unrelated to aseptic manufacturing. Cause: The category-exclusive knowledge base collection is not bound, or the similarity threshold is set too high, filtering out eligible professional documents.
- Phenomenon: In the same conversation window, subsequent knowledge base retrieval latency increases significantly, and some requests time out beyond 60 seconds. Cause: The number of documents retrieved in a single search is not limited, and redundant data from conversation context is not cleaned, resulting in excessive historical content being loaded for each retrieval.
- Phenomenon: Some questions do not trigger knowledge base retrieval, and general answers are generated directly. Cause: Forced retrieval configuration is not enabled, or historical questions in the conversation context interfere with the retrieval trigger logic.
How to confirm the configuration is correct
- Access the knowledge base retrieval configuration page of FastGPT v4.17.0, confirm that the category-exclusive knowledge base collection is bound.
- Submit a test question containing professional terms such as "aseptic filling records" and "batch release", check whether the retrieval results include documents of the corresponding category.
- Submit 3–5 related questions consecutively, observe whether the latency of each retrieval remains within a reasonable range.
- Adjust the similarity threshold to different values within the specified range, confirm that the relevance of retrieval results changes as expected.
The values provided are common starting points and should be validated against sample sets relevant to the specific deployment context.