Tool Calling and Plugins for Medical Insurance Settlement

Data Characteristics

Medical insurance settlement quality documents primarily cover various files related to healthcare service fee settlements between hospitals and medical insurance bureaus. Data sources include settlement lists exported from hospital HIS systems, patient admission records, expense details, and policy notices/audit rules issued by medical insurance bureaus. Policy notices are typically released quarterly or annually. Settlement lists are generated monthly or in batches. Document structures are complex, containing both structured table data and extensive unstructured text descriptions, such as medical record summaries and treatment process notes. Fields include patient basic information, diagnosis codes (e.g., ICD-10), surgical procedure codes, drug codes (e.g., ATC), service item codes, expense categories, and settlement amounts. Units often involve RMB currency amounts, quantities, and days.

Constraints from "Tool Calling and Plugins"

Diverse data sources and varying update frequencies for medical insurance settlement data require flexible tool calling to adapt to different data interfaces and retrieval strategies. Complex document structures mean that when calling plugins for information extraction, precise recognition of structured fields and extraction of key information from unstructured text are both necessary. For example, extracting discharge diagnosis from patient admission records requires natural language processing capabilities. Extracting self-payment ratio from expense detail tables requires table parsing capabilities. Field and unit standardization varies, requiring plugins to include data cleaning and normalization functions. An example is mapping different drug name expressions to standard ATC codes. The timeliness of medical insurance policies also places high demands on tool calling, ensuring that invoked rules and parameters are always based on the latest policy versions.

Configuration Settings

Configuration Item	Suggested Value	Rationale
`max_tokens`	`4096`	Accommodates long texts like medical insurance policy documents and patient admission records, ensuring completeness.
`temperature`	`0.1`	Medical insurance settlements involve precise facts. A low `temperature` reduces the risk of model hallucinations.
`tool_retrieval_top_k`	`5`	Ensures selection of the most appropriate tools from multiple relevant options for complex queries.
`chunk_size`	`800` characters	Balances text block size, retaining context while avoiding overly long individual blocks.
`overlap_size`	`100` characters	Ensures sufficient overlap between text blocks to handle cross-block information continuity.
`parser_timeout`	`600` seconds	Provides ample parsing time when processing large-scale settlement lists or complex policy documents.

Common Mistakes

Symptom: Settlement amounts returned after tool calling do not match actual amounts. Reason: Data cleaning failed to correctly handle decimal places or currency unit differences from various data sources, leading to incorrect parsing of the amount field.
Symptom: The model fails to identify specific policy clauses in medical insurance settlement documents, leading to incorrect judgments. Reason: The knowledge base was not updated with the latest medical insurance policy documents in a timely manner, or the text chunking strategy caused key clauses to be split.
Symptom: A Tool call Parser error occurs during the tool_call process. Reason: The tool call parameters returned by the model do not conform to expectations. An example is a missing arguments field or a type mismatch.

Verification

Select a test document containing complex policy clauses and multiple settlement records. Observe if tool calling accurately extracts all key fields (e.g., diagnosis code, settlement date, personal payment amount).
Verify if the knowledge base is synchronized with the latest policy updates from the medical insurance bureau. Test if the model can make correct judgments on simulated cases based on the new policies. Check the policy_version field.
Simulate a settlement list processing flow that includes anomalous data (e.g., empty amounts, incorrect date formats). Check if the error handling mechanism after tool calling effectively captures and reports issues. Observe the error_code or status_message fields.
Monitor the average response time of tool calls in the actual operating environment. Ensure it falls within an acceptable latency range for the business.

The values provided are common starting points. Measure against specific 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.