Data Characteristics
Pharmaceutical e-commerce platforms primarily source data from official drug manufacturer inserts, product catalogs, regulatory approval documents, and the platform's own product detail pages, user reviews, and Q&A records. This data updates frequently. Dynamic information, such as drug batches, expiration dates, inventory, and prices, may update daily or in real-time.
Drug inserts typically include standard fields like generic name, brand name, ingredients, indications, dosage, contraindications, and side effects. E-commerce detail pages add product features, packaging specifications, and promotional information. Specific data fields include numerous medical terms, drug approval numbers, production dates, expiration dates, and storage conditions. Units involve milligrams (mg), milliliters (ml), and international units (IU).
Constraints on Knowledge Base Retrieval and Recall
The dynamic nature of pharmaceutical e-commerce data requires efficient index update mechanisms in the knowledge base. This ensures the timeliness and accuracy of retrieval results, preventing outdated drug information or incorrect inventory.
The structured nature of product inserts enables precise field-based matching and semantic understanding. However, it also necessitates processing a large volume of specialized terminology and abbreviations. Fields like drug approval numbers and expiration dates are crucial for compliance and user decisions. Retrieval must accurately identify and recall this information.
User inquiries about medications demand high rigor. Retrieval results must strictly adhere to the knowledge base's scope, preventing generative answers from introducing external, uncertain information. This requires stricter recall strategies and filtering mechanisms. The presence of multiple units requires correct contextual understanding during text segmentation and embedding generation to avoid misinterpretations.
Configuration Settings
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
Chunk size (Segment Length) | 800–1200 characters (characters) | Drug inserts are content-dense. Longer segments help retain complete medical concepts and dosage information. |
Overlap Length | 100–200 characters (characters) | Ensures contextual continuity across segments, especially when describing critical information like indications and contraindications. |
Recall count (Recall Count) | 5–8 entries (items) | Given the complexity of drug information, increasing the recall count can improve coverage for subsequent re-ranking and filtering. |
Similarity threshold (Similarity Threshold) | Calibrate by actual measurement | Determine through test sets based on actual query scenarios and data characteristics to ensure high relevance in recall. |
Rerank result count (Re-ranked Return Count) | 3 entries (items) | Focuses on core information most likely needed by the user, reducing irrelevant interference and improving response efficiency. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds (seconds) | Drug inserts and similar documents can be long. Parsing requires more time. This prevents parsing failures due to timeouts. |
Common Mistakes
- RAG retrieval results include section titles unrelated to the query. This occurs when document parsing fails to effectively distinguish between body text and structured titles, leading to titles being embedded and recalled as independent text segments.
- AI answers cite information outside the knowledge base. This happens when the model lacks strict knowledge base boundary constraints. If recall results are insufficient to support an answer, the model may generalize or speculate.
- The system becomes unresponsive or errors out after uploading large product catalog files. This is due to
UPLOAD_FILE_MAX_SIZEorPARSE_FILE_TIMEOUT_SECONDSbeing configured too small, causing file upload or parsing to time out.
How to Confirm Correct Configuration
- For typical drug inquiry statements, execute retrieval and verify the recalled raw text segments. Check if they fully contain core information without redundancy.
- Examine the AI-generated answers. Confirm all cited information is traceable to specific documents and paragraphs within the knowledge base, and that no external content is generated.
- Upload and parse drug inserts of different sizes and formats (e.g., PDF, DOCX). Observe file parsing status and time taken. Ensure all files are successfully parsed and indexed.
- Simulate high-concurrency query scenarios. Monitor system resource utilization and response time. Ensure retrieval services are stable and available under actual business loads.
Note: The values provided are common starting points. Measure against specific 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.