What this category of data looks like
Data sources include approved formulation processes, MES electronic batch records, LIMS systems, quality management systems, and equipment environmental monitoring data. Daily documents cover master production files, batch production records, batch packaging records, cleaning records, inspection records, deviation CAPA change documents, and batch release approval forms. Updates occur per production batch, each critical process step, and as specified in the equipment environmental monitoring plan. Fields include material batch numbers, process parameter values, equipment operation data, test result values, deviation descriptions, and more. These cover quality units, volume units, temperature and pressure parameter units, and other relevant units.
What constraints these characteristics impose on vector models and indexing
Category data is updated by batch and critical step. Historical versions must be retained for GMP compliance traceability. Indexing must support incremental updates by batch dimension. This avoids performance loss from full reindexing. Documents contain structured parameters and unstructured maintenance descriptions. Vector models must balance semantic matching for structured fields and understanding of unstructured text. Content related to equipment inspection and maintenance often links to specific equipment models and fault codes. Indexing must support hybrid retrieval of semantics and keywords. This improves retrieval accuracy. Data involves compliance requirements. Retrieved content must be traceable to the original document batch and version. This ensures auditability for reviews.
Configuration Settings
The configuration recommendations below apply to FastGPT v4.17.0.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
| Chunk length | 800–1200 characters | Adapt to the semantic integrity of long documents such as batch records, and avoid including content spanning multiple process steps in a single chunk |
| Overlap length | 100–200 characters | Retain contextual connections between adjacent process steps, and prevent loss of equipment maintenance process logic after splitting |
| Similarity threshold | 0.72–0.85 | Differentiate between equipment fault descriptions and general process parameter descriptions, and reduce false recalls of irrelevant content |
| Single search limit | Top 8 entries | Match the data volume range of single-batch records for GMP batch record review, and avoid redundant information interfering with question-and-answer generation |
| 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 | Set from representative sample documents | defaultToken is the default token count for direct text chunking; tune it on representative source documents. See the v4.17.2 model-field reference. It is not a vector dimension, embedding-model input cap, or retrieval/LLM context. |
Three Common Misconfigurations
- Symptom: After configuring a third-party embedding model channel, the system prompts that no available channel is found. Cause: The call key for the corresponding model is not configured in the FastGPT system, or the channel address does not comply with the interface specifications of the embedding model.
- Symptom: Knowledge base search response time exceeds expectations. Stuttering occurs even with hardware configuration of 8 cores, 64GB memory, and RTX2070. Cause: The chunk length and single search limit are not adjusted according to the category data characteristics, resulting in too many retrieved documents exceeding the hardware processing capacity.
- Symptom: Vector retrieval results include a large number of non-maintenance-related process parameter records, with low matching accuracy. Cause: The similarity threshold matching the category data is not set, failing to effectively distinguish between equipment inspection content and general production parameter content.
How to Verify Successful Configuration
- Import a single batch of batch production records and equipment inspection documents, run a query related to equipment faults, and check whether the retrieved results include maintenance step content for the corresponding equipment.
- View the system index logs to confirm that only newly added batch documents are subject to incremental indexing, and no full reindexing operations are triggered.
- Check the embedding model call status to confirm that there are no consecutive call failure logs, and the returned vector data format complies with system requirements.
- Adjust the test range of the similarity threshold, and verify whether the matching accuracy of the retrieved results meets the expectations of the business scenario.
The values provided are common starting points and should be measured against the reader's own samples.