What the Data for This Category Looks Like
Data for this category primarily comes from cell banks and seed lots, process analytical systems, clean zone monitoring records, MES, LIMS, and equipment integrity test reports. Daily documents include cell bank or seed lot records, culture and purification records, aseptic filling records, environmental monitoring data, media fill simulation results, sterility test reports, and batch release documents. Data updates occur after each batch production completes, via real-time monitoring data synchronization, and per fixed monitoring plan updates. Document fields include batch identification, process parameter values, environmental monitoring indicators, test results, and some parameters have clear units such as culture temperature and sedimentation bacteria colony count.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
Data in this category is scattered but closely linked, covering cross-system records across the entire production process. Retrieval must support associated recall of same-batch process, environmental monitoring, and test data. Real-time updated clean zone monitoring data and batch records updated per plan require the retrieval pipeline to support dynamic synchronization of the latest data sources. A single document contains a large number of precise process parameters with units and compliance content. Retrieved content must accurately match clauses of process specifications and technical standards, avoiding interference from irrelevant information. Aseptic manufacturing data has high traceability requirements; retrieval results must clearly mark document sources and update times to meet compliance audit needs.
Configuration Settings
This configuration is written based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Similarity Threshold | 0.85–0.92 | Matches the precision requirement for aseptic process parameters and compliance clauses, preventing low-correlation results from interfering with compliance judgments |
| Maximum Single Recall Count | Top 8–12 | Covers the number of core documents for a single batch production, while controlling the total context length to avoid retrieval timeouts |
| Context Length Limit | 800–1200 characters | Adapts to the content length of single-batch documents, avoiding exceeding the model's context window limit |
| Reranking Switch | Enabled | Rearranges initial recall results by compliance relevance, prioritizing content directly related to process specifications and technical standards |
| Query Expansion | Enabled | Automatically expands technical term queries to cover synonyms and different expressions of compliance clauses, improving cross-document recall accuracy |
| Collection ID Configuration | Configured by business domain | Associates exclusive datasets for cell banks, production processes, clean zone monitoring, and other categories, excluding cross-domain irrelevant data interference |
Three Common Misconfigurations
- Phenomenon: After calling the knowledge base interface to create a collection, the conversation interface cannot associate the specified dataset. Cause: The collection ID parameter was not passed in the conversation configuration, or the parameter format does not meet requirements.
- Phenomenon: The total time taken for retrieval and question optimization exceeds a reasonable range. Cause: The maximum single recall count is set too high, or the lightweight reranking logic is not enabled, resulting in overly long context or excessive redundant results.
- Phenomenon: A large amount of content unrelated to process specifications appears in retrieval results. Cause: The similarity threshold is set too low, causing low-correlation results to be recalled and interfering with compliance judgments.
How to Confirm the Configuration Is Correct
- Enter the dataset management interface, verify that the business domain grouping of the collection ID configuration matches the document sources of the current category.
- Initiate a test query to confirm that the retrieval results include process parameters and compliance documents for the corresponding category, and that the result sorting meets expectations.
- Check the retrieval log to confirm that the actual values of parameters such as similarity threshold and reranking switch match the configured values.
- Test associated recall of cross-batch data to confirm that the association relationship between multi-system documents is correctly identified.
The values provided are common starting points and should be evaluated against samples specific to the deployment context.