Data Characteristics
Medical insurance access product data primarily originates from official documents published by national and local medical insurance bureaus. These include the National Medical Insurance Drug Catalog, provincial and municipal supplementary catalogs, renewal rules for negotiated drugs, payment standard documents, and related policy interpretations. Documents are typically in PDF or Word format. Content is highly structured, often containing key fields such as drug names, medical insurance payment scope, reimbursement ratios, limited payment conditions, and access dates.
Data updates frequently. The national medical insurance catalog usually adjusts annually. Local policies may release supplements or revisions periodically based on actual circumstances, making information highly time-sensitive. Documents often contain medical terminology, generic drug names, brand names, and complex descriptions of payment limitations.
Constraints on Document Parsing and Chunking
The structured nature of medical insurance access documents requires effective identification and extraction of key information during parsing. Drug names and payment conditions often appear in tables or specific paragraphs. High update frequency means the knowledge base needs to support rapid incremental updates and version management. Old policy documents should be accurately marked or replaced to avoid recalling outdated information.
Complex medical terminology and payment condition descriptions in documents demand a high granularity for chunking. Large chunks can lead to semantic confusion. Small chunks can lose context. Additionally, common tabular data in documents requires specific parsing strategies to ensure the integrity of internal table logic. For example, a drug and its corresponding reimbursement conditions must not be separated.
Configuration Parameters
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
UPLOAD_FILE_MAX_SIZE | 100 MB | Medical insurance policy documents can contain many tables and images, resulting in relatively large file sizes. |
chunk_size | 500-800 characters | Balances the completeness of medical insurance policy descriptions with knowledge point granularity, preventing truncation of key information. |
chunk_overlap | 50-100 characters | Ensures contextual coherence, especially at clause transitions, improving recall accuracy. |
parse_table_mode | true | Medical insurance documents frequently use tables to display drug information and payment conditions; table parsing must be enabled. |
doc_type_filter | PDF, DOCX | Medical insurance documents are primarily published in these two formats; restricting types improves parsing efficiency. |
max_tokens_per_chunk | 2000 | Limits the maximum number of tokens per chunk, aligning with the processing capabilities of mainstream large models. |
Common Pitfalls
- Uploading Excel spreadsheets to the knowledge base results in an error. The system does not directly support Excel files as knowledge base documents by default. Convert them to PDF or other supported text formats.
- Uploading large PDF documents leads to redundant or semantically incoherent chunking results. This happens when
chunk_sizeandchunk_overlapare not adjusted for the tables and complex structures in medical insurance documents. - After document parsing, some critical medical insurance payment limitation conditions are not accurately extracted. This occurs when
parse_table_modeis not enabled or custom delimiter settings are incorrect, causing table content to be incorrectly merged or split.
Verification
- Upload a typical medical insurance policy document (e.g., the National Medical Insurance Drug Catalog). Review the chunk preview to ensure key information like drug names and payment conditions are semantically complete within individual chunks.
- Randomly select tabular content from the document. Verify that the parsed chunks retain the internal logical relationships of the table, such as a drug and its corresponding payment ratio or limitation conditions.
- Test with query statements containing specific medical terminology or policy clauses. Observe if recall results accurately hit relevant chunks. Evaluate the contextual completeness of the recalled chunks.
The values provided are common starting points and should be measured 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.