Data Characteristics
Attenuated inactivated vaccine regulations and standard documents originate primarily from the National Medical Products Administration, World Health Organization (WHO) guidelines, and internal Quality Management System (QMS) documents from pharmaceutical companies. These documents typically exist as PDFs, Word files, or scanned images. Content covers the entire lifecycle: R&D, production, quality control, batch release, storage, transport, and post-market surveillance.
Update frequency varies. National regulations and WHO guidelines are usually revised annually or every few years. Internal SOPs update in real-time based on regulatory changes and production process improvements, potentially with weekly or even daily minor adjustments.
Document structure: Regulatory documents often have clear chapters. SOPs include flowcharts, text descriptions, tables, and attachments. Fields and units include batch number, production date, expiration date, storage temperature (Celsius), purity (%), and titer (TCID50/mL or PFU/mL). Numerical precision requirements are high.
Constraints Imposed on Knowledge Base Retrieval and Recall
The authoritative nature and high precision requirements of attenuated inactivated vaccine regulatory documents demand extremely high recall and accuracy from knowledge base retrieval. The hierarchical structure and cross-referencing within regulations and SOPs require the retrieval system to understand context. This prevents isolated fragments from becoming answers. For example, vaccine storage temperature regulations may be scattered across multiple chapters or different documents; the system must integrate this information.
Frequent local updates, especially for SOPs, necessitate efficient incremental indexing and version management mechanisms. This ensures the knowledge base always reflects the latest regulations. Tables and figures in PDFs and scanned images pose challenges for text extraction and structured processing. Failure to parse them effectively can lead to omissions of critical parameters (e.g., dosage, batch information). Additionally, multilingual guidelines (Chinese, English) require cross-language query processing.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Segment Length) | 500–800 characters (characters) | Ensures each segment contains sufficient context, avoids splitting critical clauses, and balances retrieval efficiency. |
Chunk Overlap Length (Segment Overlap Length) | 100–150 characters (characters) | Increases semantic continuity between segments, especially in regulatory articles or SOP step descriptions, improving recall. |
Recall count (Number of Retrieved Items) | Top 8–12 entries (top 8–12 items) | Considering the rigor of regulatory documents and their multi-point referencing, increasing the number of retrieved items covers potential relevant information. |
Similarity threshold (Similarity Threshold) | Calibrate by actual measurement (Calibrate by actual measurement) | Requires testing with actual queries to balance recall and accuracy, ensuring critical regulatory clauses are recalled with high priority. |
Rerank result count (Number of Reranked Items) | Top 5 entries (top 5 items) | Uses a large language model to re-sort retrieved results, ensuring the most relevant regulatory clauses are presented first. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds (seconds) | Provides ample time for text extraction and structured parsing when processing large PDFs or complex SOP documents. |
Common Pitfalls
- Query results fail to accurately cite specific regulatory articles or SOP steps, leading to vague responses. This occurs due to segment granularity being too large or too small, destroying the original semantic integrity, or improper similarity threshold settings.
- For specific parameters of a vaccine batch (e.g., expiration date), the knowledge base returns outdated or incorrect information. This happens when the knowledge base is not rebuilt or incrementally updated after document changes, leading to data version inconsistency.
- When a user asks about key quality control checkpoints in the vaccine production process, the system returns results missing important table data. This occurs when table content in PDFs or scanned images is not correctly parsed and indexed, resulting in the loss of critical structured information.
How to Verify Configuration
- Select a representative set of regulatory and SOP documents. Construct test questions based on key clauses, numerical parameters, and process steps within them. Check if the responses directly cite the original text.
- Perform update operations using different versions of SOP documents. Then query for differences between the pre-update and post-update versions. Verify if the knowledge base accurately reflects the content of the latest version.
- For PDF documents containing complex tables and flowcharts, query for structured information within them. Check if the responses include key data from tables and descriptions of process steps.
- Monitor knowledge base retrieval logs. Analyze the distribution of
doc_idorsegment_idin query results. Ensure documents from different sources are effectively retrieved.
The values provided are common starting points and should be measured against actual 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.