Data Characteristics
Preclinical safety evaluation regulations primarily originate from official guidelines and regulatory documents published by bodies such as the National Medical Products Administration (NMPA) and the International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use (ICH). Internal standard operating procedures (SOPs), testing methodologies, and report templates from various institutions also contribute to this data.
Data updates are infrequent, typically occurring annually or every few years, coinciding with regulatory revisions or technological advancements. Documents are predominantly in PDF, Word, and Excel formats. Content covers areas like animal welfare, toxicology studies, pharmacokinetic evaluations, and biosample analysis. Field values and units are highly specialized, including dosage units such as mg/kg, time units like h, and concentration units such as μg/mL, along with various toxicity grading standards and statistical indicators.
Constraints Imposed by These Characteristics on "Source and Traceability"
The authoritative and specialized nature of the data requires precise citation of specific sections or paragraphs within original documents. This ensures the compliance and credibility of responses.
Low update frequency implies stable knowledge base content. However, when regulations are updated, the system requires timely full or incremental updates. Clear traceability logic between old and new versions is essential.
Diverse document formats, especially numerous PDFs and scanned documents, demand robust document parsing and text extraction capabilities. This ensures lossless extraction of text content, particularly for tables and figure captions.
The strictness of specialized fields and units means the model must accurately reproduce numerical values and units from the original text when generating answers. This prevents misjudgments due to unit conversions or numerical discrepancies. Therefore, during citation and traceability, focus on the completeness of text snippets and their contextual relevance.
Configuration Settings
| Configuration Item | Recommended Value | Rationale for Recommendation |
|---|---|---|
Chunk size | 500 characters | Preclinical safety evaluation documents often contain long paragraphs. This length avoids semantic breaks during splitting while ensuring an appropriate amount of information per segment. |
Recall count | Top 8 entries | Given the rigor of regulatory documents, sufficient contextual information is needed to support answers and avoid missing key provisions. |
Similarity threshold | 0.75 | Ensures recalled segments are highly relevant to the query, reducing interference from irrelevant information and improving traceability accuracy. |
Rerank result count | Top 5 entries | Re-ranks recalled segments to further filter for the most relevant ones, optimizing the quality of final citations. |
maxContext | 4096 tokens | Ensures the model can process longer contextual information, especially for complex questions involving cross-references to multiple provisions. |
CHUNK_OVERLAP | 50 characters | Appropriate chunk overlap helps retain context at segment boundaries, improving the coherence of RAG recall. |
Common Pitfalls
- Citation results include irrelevant regulatory provisions. The model's answer contains unrelated content because of ineffective filtering or re-ranking of recall results.
- The cited source's document page number or section is inaccurate. Users click the citation link and find the content does not match. This can occur if the document parser incorrectly identifies the directory structure or if the segmentation logic causes a discrepancy between the cited snippet and its original location.
- The model's answer cites outdated regulations or SOP versions. The model provides regulations that do not align with the latest requirements. This happens when the knowledge base is not updated promptly or version management mechanisms fail, leading to the recall of old data.
Verification Steps
- For typical safety evaluation questions, verify the accuracy of regulatory provisions cited in the model's answer. Confirm each citation link directly navigates to the corresponding paragraph or page in the original document.
- Select an updated safety evaluation document. After updating the knowledge base, verify the model correctly cites the new version's content and no longer cites the old version's content.
- Use queries containing specialized terminology and units. Evaluate whether the model's answer accurately reproduces numerical values and units from the original text, and if the cited sources support this specialized information.
- Simulate scenarios involving regulatory conflicts or multiple references. Check if the model can reasonably integrate information from multiple sources and clearly indicate each citation's origin.
Note: The values provided are common starting points. Measure them against your own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.