Data Characteristics for This Category
DTP pharmacies prepare registration and declaration documents using data primarily from pharmaceutical manufacturers. This includes original R&D documents, clinical trial reports, pharmaceutical research reports, quality standards, instructions, and regulatory documents from drug administration departments. These documents are mostly in PDF format, with some Word documents and Excel spreadsheets. Document updates are infrequent, mainly occurring before drug listing, during post-approval changes, or re-registration. Documents have complex structures, containing extensive specialized terminology, charts, tables, and multi-level headings. Fields and units are highly specialized, such as chemical structural formulas, content units (mg/tablet, %), detection method parameters (HPLC, GC), pharmaceutical batch numbers, and expiry dates.
Constraints Imposed by These Characteristics on Document Parsing and Chunking
The specialized nature and complex structure of DTP pharmacy registration and declaration documents require the document parser to accurately identify multi-level headings, chart titles, and table contents. This maintains contextual coherence of knowledge. Embedded images and scanned documents in PDFs may make Optical Character Recognition (OCR) a critical step, affecting subsequent chunking accuracy. Multi-dimensional data in Excel files, such as test results for different drug batches, requires special handling to prevent data loss or context fragmentation. Due to extensive specialized terminology and abbreviations, chunking must consider semantic integrity to avoid cutting off critical information. Infrequent document updates mean initial parsing and chunking accuracy is crucial, as subsequent reprocessing costs are high.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Length) | 500–800 characters | Balances specialized terminology context with information volume per chunk, preventing semantic fragmentation from chunks that are too long or too short. |
Maximum Chunk Size (Maximum Chunk Size) | 1000 KB | Ensures a single knowledge chunk can accommodate document segments containing charts or complex tables. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Addresses the time required for OCR processing of large PDFs or scanned documents, preventing parsing failures due to timeouts. |
Recall count (Retrieval Count) | Top 8 entries (top 8) | Retrieval of registration and declaration documents typically requires more comprehensive information to support decision-making. |
Similarity threshold (Similarity Threshold) | 0.75 | A high threshold ensures retrieved knowledge chunks are highly relevant to the query intent, reducing noise. |
Rerank result count (Reranked Retrieval Count) | 5 entries (5 items) | Based on a high retrieval count, reranking selects the most relevant segments to improve the final response quality. |
Three Common Mistakes
- When parsing large PDF files, the system displays a "request timeout" error. This occurs because the
PARSE_FILE_TIMEOUT_SECONDSparameter was not adjusted, making the default timeout insufficient for OCR or complex structure parsing. - Retrieval results for Excel table data in the knowledge base are incomplete, for example, only returning partial batch information. This happens because the default chunking strategy does not effectively handle the multi-row and multi-column relationships in tables, leading to context loss.
- Content parsing errors or omissions occur in uploaded scanned PDF files. This is because the document parsing service is not enabled or an appropriate OCR engine is not configured, preventing text extraction from images.
How to Confirm Correct Configuration
- Randomly select multiple registration and declaration documents of different types (plain text, charts, tables). Upload them to the knowledge base. Check the number of knowledge chunks and content integrity for each document, ensuring critical information is not truncated.
- For Excel files, design queries that include cross-row and cross-column information. Verify that retrieval results completely restore the relational information from the original data.
- Upload a PDF document containing complex charts and scanned images. Check if the parsed knowledge chunks accurately identify chart titles and text from scanned images. Verify their retrievability through queries.
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.