Data Characteristics for this Category
Attenuated inactivated vaccine registration dossiers involve extensive biological, pharmaceutical, and clinical trial data. Data sources are diverse, including pharmacopoeia standards, regulatory documents from the National Medical Products Administration (NMPA), published scientific literature, preclinical study reports, clinical trial reports, and manufacturing process validation documents. These materials typically exist as PDFs, Word documents, Excel spreadsheets, and structured databases. Pharmacopoeia and regulatory documents undergo periodic revisions. Clinical trial data and manufacturing process documents are continuously generated and iterated during R&D and submission. Document internal structures are complex, containing numerous specialized terms, figures, and tables, such as strain passage records, virus titer assay results, animal protection rate data, and human immunogenicity trial reports. Fields and units are highly specialized, like biological activity units such as TCID50/mL, IU/mL, pg/mL, and key pharmaceutical information such as batch number, expiration date, and storage conditions.
Constraints on "Reference and Traceability" Due to These Characteristics
The complexity of attenuated inactivated vaccine registration dossiers imposes stringent requirements on reference and traceability. Multi-source heterogeneous data necessitates robust document parsing capabilities during knowledge base construction. This ensures accurate identification and extraction of key data from figures and tables. For example, specific values and batch information from virus titer assay reports must be precisely linked to their table locations in the original text. Periodic updates to regulatory documents require knowledge bases to have version management functionality to differentiate between historical versions of regulations. Due to the highly specialized content, models must precisely cite supporting evidence from original text passages when generating answers. This prevents information distortion caused by comprehension deviations. Any answer with unclear traceability can lead to reliability issues in the submission dossier, especially concerning critical safety and efficacy indicators like toxicity and immunogenicity. Furthermore, a large number of specialized terms and abbreviations require the model to maintain contextual consistency when citing, avoiding semantic drift.
Configuration Settings
| Configuration Item | Suggested Value | Rationale for this Value |
|---|---|---|
Chunk size | 500–800 characters | Vaccine dossier paragraphs are often long, containing multiple key pieces of information. Overly short segments can lose context; overly long segments increase recall noise. |
Recall count | 8–12 entries | Ensures coverage of relevant clauses from multiple preclinical, clinical reports, and regulatory documents, improving information comprehensiveness. |
Similarity threshold | 0.78–0.85 | This threshold helps filter out semantically weakly related references, improving precision, especially for highly specialized terms and data. |
Rerank result count | 5 entries | After initial retrieval, the reranking model further filters for references most relevant to the query intent, reducing redundancy. |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Addresses long parsing times for large PDF documents (e.g., clinical trial reports hundreds of pages long), preventing parsing timeouts. |
maxContext | 2000–3000 Tokens | Ensures the model has sufficient context to understand complex questions and accommodate multiple reference sources, particularly original regulatory text with long paragraphs. |
Three Common Mistakes
- Generated answers do not match the cited original content, leading to "answer irrelevant to citation" situations. This typically results from improper knowledge base segmentation strategies, where critical information is split, or the retrieval algorithm fails to precisely match semantically relevant passages.
- Cited sources point to an outdated document version, leading to citations of revised pharmacopoeia standards or regulatory clauses. This occurs when the knowledge base lacks effective version management or its update strategy fails to synchronize with the latest documents in a timely manner.
- When generating information about vaccine batches or experimental data, critical numerical or unit information is missing from citations. This stems from incomplete extraction of structured data within tables or figures during the document parsing phase, leading to loss of original data.
How to Confirm Proper Configuration
- Randomly select at least 20 answered questions related to vaccine submissions. Verify that key information in the answers can be precisely supported by the cited original passages.
- Check that cited sources correctly link to the specific page numbers or sections of the original documents, verifying traceability accuracy. For regulatory citations, confirm they are the currently effective versions.
- For complex questions involving tables and figures, verify that the cited content includes key numerical values and units from the tables and is fully consistent with the original content.
Note: The values provided are common starting points. Measure against your own samples to determine optimal settings.
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.