This guide is written for FastGPT v4.17.0, and provides configuration guidance for the citation sourcing and traceability step in cross-review of technical documents for active pharmaceutical ingredient and pharmaceutical intermediate production scenarios.
What data looks like for this category
Data for this category comes from registered process documents, raw material batch ledgers, DCS control systems, weighing systems, LIMS laboratory information management systems, chromatography data systems, and QMS quality management systems. Data updates trigger on every production batch and every process step. Daily documents include batch production records, material balance sheets, reaction and purification records, cleaning validation records, impurity profile reports, and deviation and change control documents. Core fields include batch number, reaction temperature, feeding amount, impurity content, and solvent residue amount. Corresponding units are degrees Celsius, kilograms, milligrams per liter, and grams per cubic meter.
What constraints these characteristics impose on the "citation sourcing and traceability" link
Data for this category updates frequently and is linked to the entire production process. Citation traceability for cross-review of technical documents must accurately match batch numbers to avoid process deviations caused by cross-batch citations. Mixed storage of multiple types of documents requires the traceability system to recall documents by scenario. For example, impurity profile reports must be prioritized for queries related to impurity testing, and reaction records must match questions about process parameters. Associated indexes must be established for homologous data from multiple systems to ensure cross-traceability between DCS records and LIMS test reports. Production compliance risks require that traceability results must point to original production and test documents, not secondarily processed summary content. Data source systems and generation times must also be clearly marked.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
quoteMaxToken | 800–1200 token | Single documents for active pharmaceutical ingredients typically range from 500 to 1500 characters. This range preserves complete citation snippets without exceeding context limits |
| Similarity threshold | 0.75–0.85 | Low-relevance general process descriptions must be filtered out to accurately match core business fields such as batches and impurities |
| Single search limit | Top 6 entries | Citations for multi-step processes must cover reaction, purification, testing and other links. 6 entries can cover associated documents for typical production workflows |
responseData | Enable original data fields | Core fields such as batch numbers and test results from source documents must be displayed to meet compliance traceability requirements |
detail | Enable detailed traceability | Data source systems and generation times must be marked to comply with compliance traceability requirements for active pharmaceutical ingredient production |
Three common mistakes
- Phenomenon: The conversation returns citations, but the corresponding documents are not uploaded to the current knowledge base. Cause: The retrieval scope is not restricted to only the currently bound local knowledge base, or previously associated external data source indexes are not cleared.
- Phenomenon: The order of citation snippets does not match the relevance of the query keywords, and low-relevance documents appear first. Cause: The re-ranking link is not configured, or the number of recalled documents exceeds a reasonable limit, causing the sorting logic to fail.
- Phenomenon: Citation displays are incomplete, only showing the beginning or end of the snippet. Cause: The value of
quoteMaxTokenis too small to cover complete key citation content.
How to confirm the configuration is correct
- Initiate a query that includes a specific batch number, and check whether the returned citation documents all correspond to this batch, with no cross-batch content.
- View the citation display area, and confirm whether fields such as source system name and generation time are included to meet compliance traceability requirements.
- Adjust query keywords to verify whether the recall priority of different types of documents (such as impurity profile reports, reaction records) matches the business scenario.
- Test turning off the citation configuration, confirm that no source annotations are displayed on the conversation interface, and verify the effective scope of the configuration.
The values provided are common starting points and should be measured against the reader's own samples.