Data Characteristics
Medical insurance access documents originate from policy files, drug catalogs, and diagnostic project catalogs published by national and provincial medical insurance bureaus. They also include guidelines from relevant industry associations. These documents update quarterly or semi-annually, with ad-hoc updates for major policy changes. Documents are primarily PDFs, containing numerous tables, nested lists, legal clauses, and medical terminology. Fields include drug names, insurance coverage, reimbursement ratios, restrictions, indications, and contraindications. Units involve currency (yuan), percentages (%), time (months/years), and measurements (mg/ml).
Constraints Imposed by These Characteristics on "Document Parsing and Chunking"
The semi-structured nature of medical insurance access documents demands advanced parsing capabilities. Extensive tables and nested lists mean simple text segmentation can break information integrity, losing critical associations. For example, drug reimbursement ratios often appear in the same row or adjacent cells as specific indications or usage conditions. The specialized and sometimes ambiguous nature of medical terminology requires the parser to understand semantics to avoid misinterpretations. Unpredictable update frequencies necessitate a parsing process that supports rapid response and incremental updates to maintain knowledge base timeliness. The mix of fields and units challenges entity recognition and standardization, requiring accurate identification of the actual meaning behind numbers, such as distinguishing "500 mg" from "500 yuan."
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Length) | 800–1200 characters | This length range effectively preserves complete information for individual policy entries, considering the integrity and contextual relevance of medical insurance policy texts, preventing critical information truncation. |
Chunk Overlap Length (Overlap Length) | 150 characters | Ensures sufficient contextual overlap between adjacent chunks. This helps capture cross-chunk semantic relationships during retrieval, especially at table row boundaries or within complex lists. |
File Type Whitelist | pdf, docx | Medical insurance access documents are primarily published in PDF and DOCX formats. Limiting file types improves parsing efficiency and avoids processing irrelevant formats. |
Table Parsing Mode (Table Parsing Mode) | Structured Parsing | Medical policy documents contain extensive tabular data. Structured parsing accurately extracts row and column information from tables, maintaining data integrity, such as the correspondence between drug names, reimbursement scopes, and restriction conditions. |
Entity Recognition Enabled | Yes | Enabling entity recognition accurately identifies and annotates key medical and economic entities in documents, such as drug names, disease names, amounts, and percentages. This provides more precise semantic information for subsequent retrieval and question answering. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Medical insurance policy files can be lengthy, especially with complex tables and images. Increasing the timeout prevents parsing failures due to excessive processing time, ensuring large documents are successfully handled. |
Three Common Mistakes
- Symptom: The knowledge base lacks certain drug reimbursement scopes or restrictions, leading to incomplete answers during retrieval. Reason: Table structured parsing was not enabled during document parsing, or
Chunk size(Chunk Length) was too short, causing table row data to be incorrectly split. - Symptom: The API returns a parsing status of "parsing in progress" for an extended period, potentially resulting in a
parsing failederror. Reason: ThePARSE_FILE_TIMEOUT_SECONDSparameter is set too low. For PDF documents with many complex tables or high-resolution images, parsing time exceeds the preset threshold. - Symptom: The knowledge base contains numerous duplicate policy clauses, leading to redundant retrieval results. Reason: The document preprocessing stage did not effectively deduplicate different document versions, or
Chunk Overlap Length(Overlap Length) was too large, causing duplicate information to be indexed multiple times.
How to Confirm Proper Configuration
- Upload representative, complex medical insurance policy PDF documents. Examine the content of the parsed knowledge chunks to verify the completeness and accuracy of table data, nested lists, and legal clauses.
- Upload large medical insurance policy files via the API. Monitor the parsing status to ensure files transition smoothly from "parsing in progress" to "ready" without timeout errors. This confirms the
PARSE_FILE_TIMEOUT_SECONDSparameter is appropriate. - Randomly select multiple knowledge chunks from the knowledge base. Verify their source documents to confirm reasonable segmentation and preserved semantic integrity, avoiding truncation of critical information.
- Use queries containing specific medical terminology and numerical values to test the knowledge base's retrieval capabilities. Confirm entity recognition functions correctly and accurately matches knowledge chunks containing relevant fields and units.
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.