What Data for This Category Looks Like
Data for this category comes from approved formulation processes, material batch information, MES electronic batch records, equipment and environmental monitoring data, LIMS systems, and quality management systems. Daily generated documents include master production documents, batch production records, batch packaging records, cleaning records, inspection records, deviation/CAPA/change documents, and batch release review forms. Updates occur each production batch, upon completion of each key process step, or per a fixed monitoring plan. Document structures follow the full batch execution workflow, including fields such as material weighing, process parameters, inspection results, and deviation handling. Some fields require statutory units of measurement.
Constraints on Citation Sources and Traceability
Data sources are scattered and tied to unique batch identifiers. Citation and traceability workflows must accurately bind full workflow data for a single batch to avoid cross-batch confusion. Multiple document types with fixed structures require support for targeted recall of associated content by document category. Update frequency is high and tightly linked to production steps, so referenced content must sync with the execution node of the current production batch. Batch records under GMP requirements must be fully traceable. Citation and traceability results must clearly link document generation time and approval status, and cannot omit verification data from key links.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the core content length of a single batch record page, avoids splitting critical process information across paragraphs |
| Maximum Context | Top 10 recall results | Covers core associated documents for a single batch production, avoids redundant data interfering with review |
quoteMaxToken | 1500 token | Limits the output length of a single citation, ensures complete display of key fields in batch records |
| Similarity Threshold | 0.75–0.85 | Matches the professional term similarity of pharmaceutical production documents, filters irrelevant general process descriptions |
| Single Search Limit | Top 6 candidate segments | Balances recall precision and query efficiency, adapts to the needs of multi-document cross-review |
| Reranking | Enabled | Prioritizes recall of document segments most closely linked to the current batch, aligns with GMP traceability priority requirements |
Three Common Configuration Errors
- Citation snippets are not rendered in Markdown format. Cause: The format parsing switch for knowledge base documents is not enabled. Split raw text displays directly, losing typesetting structure.
- Associated knowledge base documents cannot be recalled during querying. Cause: Chunk length is set too large, truncating critical process segments so they cannot match query semantics. Or the similarity threshold is set too high, filtering valid recall results.
- The citation display function cannot be disabled. Cause: The
responseDataparameter is incorrectly configured to force citation returns, without adjusting to the mode that only returns response content.
How to Confirm Proper Configuration
- Upload batch production records and inspection records for a single batch. Initiate a query related to process review. Check that citation snippets in returned results use correct Markdown typesetting.
- Enter a material batch number for a specific batch. Verify that all recalled reference documents are bound to that batch identifier, with no cross-batch irrelevant content.
- Access the FastGPT v4.17.0 configuration panel. Confirm the
detailparameter enables display of document traceability fields, including document generation time and approval status. - Test disabling the citation function. Confirm response content no longer includes links or snippets of citation sources, meeting business requirements.
The values provided are common starting points and should be measured against the reader's own samples.