Knowledge Base Retrieval and Recall for Medical Insurance

Data Characteristics for Medical Insurance Access Documents

Medical insurance access R&D documents originate from various sources: national and provincial medical insurance bureaus' policy documents, drug catalog adjustment notices, pharmaceutical companies' pharmacoeconomic evaluation reports, clinical trial data, and internal compliance records. These documents update frequently, especially during annual medical insurance catalog adjustments. Document formats vary, including policy texts in PDF, application templates in Word, and data attachments in Excel. Structurally, policy documents often contain introductions, specific clauses, and appendix lists. Clauses include key fields like drug names, indications, payment scope, restricted payment conditions, and negotiated prices. Data fields are highly specialized, for example, ATC classification codes, ICD-10 disease codes, and national medical insurance drug codes. Units include milligrams, milliliters, yuan, and percentages. It is crucial to distinguish between generic and brand names.

Constraints on Knowledge Base Retrieval and Recall

The authoritative nature and update frequency of medical insurance access documents require the knowledge base to support efficient document synchronization and version management. This ensures timely and accurate retrieval results. The large number of specialized terms and codes in documents demands that tokenizers and embedding models accurately recognize and vectorize them to avoid semantic loss. Diverse document structures and formats, especially tabular data, challenge the parser's ability to extract structured information. This ensures key fields like payment scope and restricted payment conditions are correctly identified and indexed. The complexity and interconnectedness of medical insurance policies mean that a single keyword search is often insufficient. More intelligent semantic retrieval and multi-hop query capabilities are necessary. Numerical information, such as amounts and percentages, requires support for range queries and unit conversions during retrieval to meet compliance analysis needs.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`Segment Length`	`800–1200 characters`	Medical insurance policy clauses are often lengthy. This length helps preserve complete semantic context and reduces the risk of truncating critical information.
`Overlap Length`	`100 characters`	Ensures sufficient overlap between adjacent text segments. This covers related information spanning paragraphs, especially when processing continuous policy texts.
`Recall Count`	`Top 8–12 items`	Medical insurance access queries often require synthesizing multiple policies or reports. Increasing the recall count improves comprehensiveness.
`Similarity Threshold`	`0.75–0.82`	The medical insurance domain demands high retrieval accuracy. This threshold ensures relevance while filtering out irrelevant or weakly related results.
`Max Concurrent File Processing`	`5`	Medical insurance policy documents are typically large and complex to parse. Limiting concurrency prevents system overload and ensures parsing stability.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	For PDF files with complex tables or extensive text, extending the parsing timeout ensures complete file processing.

Common Pitfalls

Uploading documents with Chinese filenames may result in garbled characters or parsing failures in the knowledge base system. This occurs due to incompatible encoding configurations between the system and the filenames.
Retrieval results may contain many irrelevant or low-quality document snippets. This happens when the Similarity Threshold is set too low, failing to effectively filter out noise.
When querying numerical information, such as medical reimbursement ratios, retrieval results may not accurately display or support range queries. This usually occurs because numerical fields were not structurally extracted during document parsing and were treated as plain text.

Verification Steps

Upload a batch of medical insurance access documents with Chinese filenames and various formats (PDF, Word, Excel). Verify that all files parse and ingest successfully, and filenames display correctly.
Perform semantic searches for specific drug names, indications, and payment conditions within medical insurance policies. Check if the Recall Count is sufficient and if the top results are highly relevant to the query intent.
Use queries involving numerical ranges (e.g., "drugs with reimbursement ratios above 70%"). Verify that the system returns correct policy entries and accurately identifies and displays numerical information in the document snippets.

The values provided are common starting points. Measure them 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.