Data Characteristics in this Category
Patient Assistance Program (PAP) quality documents include project plans, patient recruitment criteria, drug management guidelines, follow-up records, adverse event reports, and compliance audit reports. Data sources typically originate from pharmaceutical companies' internal project management systems, patient information exported from hospital HIS systems, and operational data from third-party service providers. Document update frequency depends on project cycles, policy changes, and audit requirements. For example, project plans may be revised annually, while patient follow-up records update in real-time. Document structures primarily consist of structured and semi-structured data, such as drug batch numbers, patient IDs, and medication dosages, which have defined fields. Unstructured data mainly appears in physician handwritten medical summaries or patient feedback records. Fields and units are specialized for the biomedical domain, including drug generic names, batch numbers, production dates, expiration dates, administration routes, dosage units (mg, ml), and adverse reaction codes (MedDRA).
Constraints Imposed by these Characteristics on "Reference Sources and Traceability"
The diversity and update frequency of PAP document data impose specific requirements on reference sources and traceability. Real-time updated follow-up records and adverse event reports require high-frequency data synchronization and indexing to ensure the timeliness and accuracy of cited content. Semi-structured and unstructured data necessitate robust text parsing and entity recognition capabilities within the knowledge base to extract key information from complex descriptions, such as identifying drug names and dosages from physician handwritten notes. Specialized fields and units mean the knowledge base must support custom dictionaries and named entity recognition to prevent citation errors due to insufficient domain knowledge. Furthermore, the rigor of compliance audit reports demands that citation results trace back to specific paragraphs in original documents for audit verification, emphasizing precise matching and localization of cited content with the original text. Document interdependencies, such as the link between project plans and recruitment criteria, require the knowledge base to establish effective inter-document links to support cross-document citation traceability.
Configuration Settings
| Configuration Item | Suggested Value | Rationale for this Value |
|---|---|---|
Chunk size (Chunk Length) | 800–1200 characters | Balances common paragraph lengths in PAP documents with semantic completeness, preventing information loss due to splitting. |
Recall count (Recall Count) | 8–12 items | Balances recall efficiency with relevance, covering multiple key information points in patient assistance documents. |
Similarity threshold (Similarity Threshold) | 0.78–0.85 | Ensures recall results are highly relevant to patient assistance queries, filtering out low-relevance content. |
Rerank result count (Rerank Return Count) | Top 5 items | Refines the content presented to the user, prioritizing the most relevant PAP details. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Accommodates the complex parsing requirements of large patient assistance project plans or audit reports. |
MaxTokens | 4096 | Satisfies the context length needed for citing detailed technical descriptions and compliance terms in PAP documents. |
Three Common Mistakes
- Uploading many small files leads to training failures for some files, preventing the knowledge base from finding corresponding content during citation. This occurs due to insufficient concurrent processing capability or single-file parsing timeouts.
- Selecting "Variable Reference" to specify a large model in the AI chat component prevents setting parameters like temperature, leading to a lack of flexibility in model output. This is due to interface logic design limitations, which do not provide an entry point for advanced parameter configuration in "Variable Reference" mode.
- The
{{id}}field in the knowledge base citation content template is empty, making it impossible to trace back to the specific data source. This happens when the knowledge base indexing fails to correctly associate or store the unique identifier of the original document.
How to Verify Correct Configuration
- Upload a patient assistance project plan containing multiple professional terms and specific values. Check if the knowledge base indexing is successful. Then, perform keyword queries to verify that the recalled content includes key information.
- Train the knowledge base with an audit report containing complex tables and multi-level headings. Then, use simulated questions to confirm that citation results accurately point to specific paragraphs in the original document.
- Query using specific fields like patient ID and drug batch number. Verify that the knowledge base precisely matches and cites corresponding patient follow-up records or drug management guidelines.
- Compare knowledge base citation results with original documents. Verify the completeness and accuracy of the cited content. Check if
{{id}}or other identifiers can trace back to the original file or data record.
Note: The values provided are common starting points. Measure them against your own samples for optimal performance.
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.