What Data for This Category Looks Like
Daily operational documents for this category include cell bank or seed lot records, culture and purification records, aseptic filling records, environmental monitoring data, media fill simulation results, sterility test reports, batch release documentation, and more. Data sources cover cell banks and seed lots, process analysis systems, clean zone monitoring systems, MES, LIMS, and equipment integrity test results. Data updates follow per-production-batch, real-time monitoring, and fixed monitoring plan workflows. Individual documents include fields such as batch number, operation time, process parameters, test results, and compliance status. Parameters typically carry clear units, such as culture temperature 37℃, pH value 7.2.
Constraints Imposed on Vector Models and Indexing
Data for this category has high update frequency and strong ties to production batches. Indexes must support incremental updates to adapt to real-time or batch-based synchronization requirements. Documents contain extensive professional process parameters and compliance descriptions. Chunking must retain the integrity of parameters and associated explanations to avoid splitting critical information. Data has extremely high compliance requirements. Retrieval results must accurately match the target batch and process link, without introducing content from unrelated batches. Additionally, individual documents have many professional fields, so indexes must support metadata association to quickly filter non-target data.
How to Set Configurations
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Documents for this category contain professional parameters and long sentence explanations. This length can retain the integrity of parameters and associated constraints, avoiding splitting critical process logic |
| Overlap Length | 100–150 characters | Retains professional terminology and contextual association between adjacent chunks, preventing loss of cross-chunk parameter associations |
| Similarity Threshold | 0.75–0.85 | Accurate recall of content related to aseptic control and process parameters is required, avoiding introduction of documents from unrelated batches or non-target links |
| Single Search Limit | Top 3 results | Key steps and parameters of process specifications are concentrated in a small number of documents. Excessive recall will introduce redundant information |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
defaultToken | Calibrated via actual testing | Token processing logic varies across embedding models, and values must be adjusted based on the actually deployed model |
Three Common Mistakes
- Files remain in the "indexing" state for extended periods after upload, with no clear error logs. The root cause is incorrect configuration of the embedding model deployment address or call permissions, which prevents indexing tasks from completing.
- Retrieval results return process specification content from multiple unrelated production batches. The root cause is failure to bind batch metadata fields during the indexing phase, which means non-target batch documents are not filtered during recall.
- Retrieved process parameters are incomplete, with missing associated operating conditions or units. The root cause is an overly small chunk length setting, which splits complete text blocks containing parameters, units, and constraint conditions.
How to Confirm Proper Configuration
- Review indexing task logs in the FastGPT v4.17.0 backend to confirm successful embedding model calls, with no timeout or error messages.
- Upload a test aseptic filling record document, run a retrieval for "media fill simulation acceptance criteria", and confirm the retrieval results include the corresponding document fragments.
- Check post-indexing document metadata to confirm that fields such as batch number and operation time are correctly associated with retrieval results.
- Adjust the similarity threshold to verify that the number of retrieval results matches the filtering needs of the target business scenario.
The values provided are common starting points and should be measured against the reader's own samples.