Data Characteristics for this Category
Patient Assistance Program (PAP) policy data originates from official documents, program manuals, application guides, eligibility criteria, and drug lists published by pharmaceutical companies or foundations. These documents are typically in PDF, Word, or structured text formats (e.g., JSON, XML). PAP policies may undergo major annual updates or minor revisions triggered by new drug approvals or changes in medical insurance policies. Document structures often include legal clauses, medical terminology, approval processes, financial terms, and pharmaceutical information. Specific fields and units include generic and brand names, dosages, specifications, and prices of drugs. Patient data includes diagnostic codes (e.g., ICD-10), treatment plans, and financial proofs (income statements, poverty certificates). Price units are typically "CNY/box" or "CNY/vial," and measurement units include "mg" and "ml."
Constraints Imposed by Data Characteristics on Vector Models and Indexing
PAP policy data often combines long texts and structured content. This requires vector models to effectively capture specialized terminology and logical relationships within the text. The document update frequency dictates the indexing rebuild cycle: major annual updates require full rebuilds, while minor revisions can use incremental updates. Complex document structures necessitate fine-grained text chunking strategies to prevent information loss or context breaks. For example, a complete application process might span multiple sections; chunking must preserve the integrity of the process. The specificity of fields and units, especially drug information and medical codes, demands higher semantic similarity understanding from vector models, as ordinary text similarity may not be accurate enough. This requires preprocessing specific fields or using domain-specific embedding models before vectorization to improve recall accuracy.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale for Recommendation |
|---|---|---|
Chunk size (Chunk Length) | 800–1200 characters | Ensures each chunk contains a complete clause or process step, preventing semantic fragmentation. |
Chunk Overlap Length (Chunk Overlap Length) | 100–200 characters | Maintains contextual continuity, especially at clause transitions, reducing information loss risk. |
Recall count (Recall Count) | top 10–15 items | PAP policy queries often require synthesizing multiple related regulations to ensure comprehensive coverage. |
Similarity threshold (Similarity Threshold) | Calibrate by testing | Dynamically adjust based on actual query results and recall accuracy to ensure result relevance. |
Rerank result count (Rerank Return Count) | top 5 items | Further improves the ranking of the most relevant results using a reranking model after initial recall. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Provides sufficient parsing time for large PDF or Word documents, preventing timeout failures. |
Three Common Mistakes
- A
{"error":{"code":"Invalid ｜ ...error in the logs usually indicates incorrect multimodal embedding model configuration or insufficient API key permissions. - Knowledge base query results missing critical clauses or drug information occur when text chunking is too fine or too coarse, leading to important context being cut off or obscured.
- After a version upgrade, old vector library data may become unusable, reporting a
400 status code no bodyerror. This typically happens when the new version uses different vector models or index formats, requiring data migration or index reconstruction.
How to Confirm Proper Configuration
- Query core patient assistance policy clauses. Check if recall results include all relevant details and if key fields (e.g., drug names, application conditions) are accurately matched.
- Simulate real patient questions, such as "What are the conditions for applying for drug XX?" Verify the completeness and accuracy of the returned results to determine if they effectively answer the question.
- Select a batch of queries containing medical terminology and specific units. Verify if the system correctly understands and recalls document snippets containing this information, for example, "How many milligrams per vial of drug XX?"
- Check the execution logs of index reconstruction tasks. Confirm that all documents were successfully parsed and vectorized, with no
PARSE_FILE_TIMEOUT_SECONDSor other parsing errors.
The values provided are common starting points. Measure them 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.