Reference Tracing for Structured Document Parsing in Smart

Data Characteristics in This Domain

Smart triage data primarily originates from clinical pathway documents, disease treatment guidelines, drug inserts, medical research reports, and medical knowledge graphs within healthcare institutions. Document update frequencies vary. Clinical pathways and treatment guidelines typically undergo quarterly or annual revisions based on the latest medical research and clinical practice. Drug inserts remain relatively stable after drug approval, updating only for adverse reactions or changes in indications. Most documents are unstructured or semi-structured text, containing extensive medical terminology, abbreviations, examination indicators, and dosage units. For example, treatment guidelines often present information in chapters, sub-sections, and tables, covering diagnostic criteria, treatment plans, and prognosis evaluations. These documents include specialized fields and units such as ICD-10 disease codes, ATC drug classification codes, mmol/L, and mg/kg.

Constraints Imposed by These Characteristics on "Reference Tracing"

The unstructured nature of smart triage data requires document parsing to accurately identify and extract key information, such as disease names, symptom descriptions, examination results, treatment recommendations, and drug dosages. The specialized and complex nature of medical terminology, along with potential synonyms and near-synonyms, demands advanced semantic understanding and information matching. Given the strictness of treatment protocols, the accuracy of reference tracing is critical. Any miscitation or missing reference can lead to inaccurate triage advice, impacting patient safety. Document update cycles mean knowledge bases require regular synchronization with the latest content, ensuring references point to the currently valid version. The presence of specialized fields and units necessitates retaining their original form and context after structured parsing. This avoids ambiguity or loss of critical information when referenced. For example, an HbA1c value of 7.0% differs significantly in meaning from 7.0 g/dL.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale for Recommendation
`Chunk size` (Segment Length)	`500–800 characters`	Balances semantic completeness with recall accuracy. Avoids segments that are too long (information overload) or too short (missing context).
`Recall count` (Number of Retrieved Items)	`Top 8–12 items`	Smart triage requires comprehensive information. Increasing recall covers more potentially relevant information.
`Similarity threshold` (Similarity Threshold)	Calibrate based on actual measurements	Balances recall and precision. Avoids citing irrelevant content while ensuring highly relevant documents are retrieved.
`Rerank result count` (Number of Reranked Items)	`Top 5 items`	Ensures the final references presented to the user are the most relevant core evidence, reducing redundancy.
`maxContext`	`4000 tokens`	Accommodates sufficient retrieved content and user queries, meeting the contextual needs of complex medical questions.
`PARSE_FILE_TIMEOUT_SECONDS`	`300 seconds`	Handles the parsing time for large treatment guidelines or medical reports, preventing parsing failures due to timeouts.

Common Pitfalls

Cited files returned by the knowledge base do not match the actual answer content. Investigation revealed that the Chunk size (Segment Length) was too long during document parsing, causing a single segment to contain multiple unrelated topics, leading to incorrect retrieval.
The system returned cited content but did not display the specific source document link. Inspection showed that the datasetId variable format was incorrectly configured, preventing proper association with the original file.
Ambiguity arose in citing specific medical terms or examination indicators. Logs indicated that the Similarity threshold (Similarity Threshold) was set too high, causing semantically similar but distinct content to be overlooked.

How to Confirm Correct Configuration

For typical smart triage questions, test whether the cited documents returned by the system cover all key information points in the answer. Check if the cited document versions are current.
Randomly sample system-generated triage advice. Verify that the specific text segments cited accurately originate from the source document. Confirm that the cited datasetId correctly links to the corresponding file.
Input queries containing specialized medical abbreviations and units. Observe whether the system correctly identifies and cites documents containing these details. Evaluate whether the cited content retains the integrity of original fields and units.

Note: The values provided are common starting points. Measure 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.