Data Characteristics
Medical insurance settlement regulations primarily originate from policy documents, implementation rules, operational guidelines, payment standards, drug catalogs, and treatment item catalogs issued by national and local medical insurance bureaus. These documents are typically published in PDF, Word, or HTML formats and are updated frequently. Some policy documents are revised annually or have supplementary explanations issued.
Document structures often include titles, chapters, clauses, and appendices. Content involves extensive specialized terminology, cost codes, settlement rules, and exceptions. Fields may include cost types, reimbursement ratios, deductibles, caps, coverage scope, and applicable diseases. Units involve monetary amounts (Yuan), percentages (%), and time (days, months).
Constraints on Vector Models and Indexing
The high update frequency of medical insurance settlement regulations requires vector models to support rapid incremental indexing, avoiding lengthy full re-indexing.
The complex hierarchical structure and extensive specialized terminology in policy documents make simple text segmentation insufficient for capturing semantic relationships. This necessitates more refined text preprocessing and chunking strategies. For example, an explanation for a clause might be dispersed across multiple paragraphs or even different appendices.
Precise matching requirements for key fields like cost codes and reimbursement ratios challenge the semantic understanding capabilities of vector models. Models must distinguish different meanings under similar descriptions. Numerical information, such as amounts and percentages, must retain its magnitude and range relationships after vectorization to support filtering and retrieval based on numerical ranges.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunk_size | 500–800 characters | Balances semantic completeness with indexing efficiency. Avoids overly long chunks that dilute core information or overly short ones that lose context. |
chunk_overlap | 100–150 characters | Ensures semantic continuity at chunk boundaries, especially when clause explanations span multiple paragraphs. |
similarity_threshold | 0.75–0.85 | Medical insurance policies require high precision. This range helps filter low-relevance results while retaining some flexibility. |
retrieve_top_k | 10–15 items | Ensures coverage of relevant clauses from multiple angles for complex queries, improving recall. |
rerank_top_n | 3–5 items | Selects the most relevant and representative clauses after reranking, improving the quality of the final answer. |
embedding_model_name | bge-large-zh-v1.5 | For Chinese medical insurance policy texts, this model performs balanced semantic understanding and similarity calculation. |
Common Pitfalls
- Knowledge base status shows "indexing" for an extended period with no progress: This often occurs when documents contain numerous images or complex tables, leading to file parsing timeouts. Check the
PARSE_FILE_TIMEOUT_SECONDSparameter setting. - After switching vector models, similarity scores show abnormal values like
10000+: This indicates that the new vector model's vector distance calculation method does not align with the system's default expectations. Adjust thesimilarity_thresholdcalculation logic or normalization process. - Inaccurate or missing results when querying specific medical insurance cost codes or reimbursement ratios: This might be due to an improper chunking strategy, where critical numerical information is truncated or disconnected from its description. Optimize text preprocessing rules.
Verification
- Upload typical medical insurance policy documents. Observe if the knowledge base indexing completes normally and check for any parsing error logs.
- Conduct multiple test queries for specific medical insurance settlement questions, such as "What is the deductible for a certain type of disease?" Evaluate the accuracy and completeness of the returned results.
- Randomly select key clauses from documents. Query their core content to verify if the recall results include the clause and its context. Check if the
similarity_thresholdeffectively filters highly relevant content. - Regularly track medical insurance policy updates. Upload newly released documents to confirm the efficiency of incremental indexing and its impact on the existing knowledge base.
The values provided are common starting points and should be measured against the reader's 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.